20 Novembre 2026

Psicologia Punto One | Psicologia e Psicologi nel Web

Menu

Menu

SCHEDA SCIENTIFICA2026Altra ricerca

Information leakage and performance overestimation in EEG-based schizophrenia detection: evidence from literature and empirical analyses.

Frigyes Samuel Racz, Gabor Csukly

Translational psychiatry · 2026

Abstract originale

Detecting schizophrenia (SZ) from electroencephalography (EEG) signals using machine- and deep learning models gained traction lately due to potential utility in early disease detection and differential diagnosis. Classification performance reports in the range of 95% accuracy and above are common; however, review of state-of-the-art literature indicates that ~65% of published works involve erroneous practices in the evaluation pipeline such as epoch- instead of subject-based data splitting or ranking and selecting features before data partitioning. The consequent information leakage can result in an overestimation of SZ detection performance. Here we explicitly test this on three open SZ-EEG datasets using gold standard classification approaches in leaky and leakage-free implementations. Results indicate that information leakage can inflate SZ classification accuracy by up to ~30%. Accordingly, best practices regarding EEG-based SZ detection must be established and promoted before this technology can be further developed into a clinical decision-making tool.

Parole chiave

HumansSchizophreniaElectroencephalography
Leggi il testo integrale originale

Testo estratto dalla pubblicazione: sono esclusi immagini, media, supplementi e parti soggette a diritti distinti.

Information leakage and performance overestimation in EEG-based schizophrenia detection: evidence from literature and empirical analyses

Detecting schizophrenia (SZ) from electroencephalography (EEG) signals using machine- and deep learning models gained traction lately due to potential utility in early disease detection and differential diagnosis. Classification performance reports in the range of 95% accuracy and above are common; however, review of state-of-the-art literature indicates that ~65% of published works involve erroneous practices in the evaluation pipeline such as epoch- instead of subject-based data splitting or ranking and selecting features before data partitioning. The consequent information leakage can result in an overestimation of SZ detection performance. Here we explicitly test this on three open SZ-EEG datasets using gold standard classification approaches in leaky and leakage-free implementations. Results indicate that information leakage can inflate SZ classification accuracy by up to ~30%. Accordingly, best practices regarding EEG-based SZ detection must be established and promoted before this technology can be further developed into a clinical decision-making tool.

Introduction

Schizophrenia (SZ) is a severe, disabling mental disorder affecting about 1% of the population [1]; however, its exact etiology and pathological mechanisms are yet unknown [2]. Since no clinically valid biomarkers are established for SZ, the disorder is currently diagnosed using extensive psychiatric interviews (DSM-5), which can be susceptible to variability arising from heterogeneity in clinical presentation and physician subjectivity. Consequently, there is immense research effort for translating SZ diagnosis, monitoring and management from perception-based practices to those grounded in biology and pathophysiology [3]. Techniques to image or monitor neural activity such as functional magnetic resonance imaging (fMRI) or electroencephalography (EEG) appear as intuitive choices for potential biomarkers in psychiatry; however, as Etkin and Mathalon [4] state in their recent review, the “low reliability and difficulty standardizing collection are the principal barriers to fMRI, along with the need to demonstrate that its superior spatial resolution over EEG and ability to image subcortical regions directly provide unique clinical value”. On the other hand, resting-state EEG markers, such as individual alpha peak frequency and power, or microstates exhibit excellent test-retest reliability [5]. Furthermore, with the emergence and increased accessibility of machine learning (ML) and especially its specialized subset deep learning (DL), automated detection of SZ and its characteristic patterns from EEG garnered particular interest lately [6, 7]. Many of these works report stellar SZ detection accuracies ranging 95%, with some even surpassing 99% [8–13]. This performance is well over current estimates of SZ diagnostic accuracy via standard clinical procedures [14], and thus could imply that computer-aided diagnostics (CAD) for SZ are within reach in the near or foreseeable future, as suggested by some reports [15, 16]. In this present work we argue that such cautious optimism is yet uncalled for, as most reported performance metrics are likely massively overestimated due to information leakage.

Leakage refers to the scenario when information from outside the training set (and in particular, from the prospective test set) is involved in the model training procedure [17, 18]. This generally results in performance overestimation and poor subsequent generalization. The latter, however, remains unconfirmed if the model is never tested on previously unseen data, which is most often the case for EEG-based classification studies working with relatively small datasets or confined data repositories. While the employed ML/DL methodologies among related works shows great variability, approaches can be sorted into two main categories: i) DL architectures utilizing native EEG or that transformed into ‘image-like’ representations (such as spectrograms or scalograms) as input, which predominantly learn group-characteristic patterns from abundant training data and ii) DL/ML frameworks utilizing ‘handcrafted’ features (such as spectral, non-linear dynamic or connectivity metrics) that attempt to capture relevant patterns of EEG activity based on a priori hypotheses [6]. Both approaches, but especially approach i) relies on vast amounts of data, which is commonly achieved by segmenting continuous EEG into smaller epochs, substantially augmenting the available sample size. On the other hand, feature extraction in ii) often results in hundreds of feature candidates and necessitates selection to reduce dimensionality and eliminate noise (i.e., irrelevant features). Consequently, two forms of leakage is especially common when working with EEG, stemming from strong inter-individual variability and high dimensionality: leakage related to data partitioning [19] and feature selection [20] (DP and FS, respectively, see Fig. 1). The first case manifests when the DP is not subject-based (i.e., where different samples from the same subject appear in both training and evaluation sets) as the model can pick up on subject-specific patterns (e.g., biometrics), while the second when the complete dataset is used to select most-discriminative features before partitioning. Despite being the leading source of error in ML and DL applications [21], leakage is only gaining traction lately in the neuroscience literature, as demonstrated in neuroimaging-based detection of post-traumatic stress disorder (PTSD) [22], Parkinson’s Disease [23] or Alzheimer’s Disease [24], or in connectome-based ML applications [25]. It was indicated that DP- and FS-related leakage has the most impact on classification performance when working with neuroimaging data [25]. Furthermore, Brookshire and coworkers [26] also reported that over 70% of translational EEG-based DL studies contained some form of leakage, likely inflating performance reports in detecting various clinical conditions. Recognizing these widely present issues prompted a recent urge for establishing best practices in similar approaches aiming for CAD using EEG [19].

Addressing the potential effect of leakage in EEG-based CAD approaches for SZ is not only relevant but pressing as well. Latest systematic reviews on the topic, [6, 7, 27, 28] indicate that the number of studies using ML or DL techniques to detect SZ from EEG grows exponentially. While over 45 articles report classification accuracies of 95% or higher for SZ (against healthy controls), none of these contemporary review articles address or consider potential leakage effects, nor has this issue been systematically investigated in the case of SZ detection. Overall, developing CAD systems for SZ is indeed important as these efforts can help identifying characteristic biomarkers of the disease, support early diagnosis and enhance differential diagnosis, all being current therapeutic challenges regarding SZ [29]. However, a potential abundance of inadequate ML/DL practices coupled with corresponding inflated performance metrics poses a real risk of misdirecting future research efforts and resources towards unfeasible or unviable solutions [18]. Therefore, addressing the issue of information leakage in state-of-the-art literature on EEG-based CADs for SZ is of utmost importance. To this end, we first screened current literature to assess the presence of DP- and FS-related leakage in previously published studies, then performed empirical testing of ‘leaky’ vs. ‘leakage-free’ classification pipelines on three open EEG datasets. We carried out these evaluations using a variety of approaches including conventional ML, DL, and state-of-the-art, transformer-based architectures. Our findings explicitly demonstrate that bias introduced by various leakage scenarios can inflate SZ classification performance by up to > 30%. To the best of our knowledge, this work presents the first synthesis of leakage in relevant recent literature, as well as the systematic assessment of leakage effects on EEG-based SZ detection evaluated via multiple approaches and on multiple real-world datasets. Furthermore, at the end of this article we propose best practices based on conclusions we derive from previous literature and empirical analyses, including a checklist provided as downloadable content to facilitate transparency and consistency among future studies.

Results

Leakage trends in current schizophrenia literature

First, we synthesized all articles collected by recent systematic reviews and broad surveys from 2023–2025 on EEG- and ML/DL-based SZ detection [6, 7, 27, 28, 30] to gain initial insight on the presence of leakage in current literature (see Supplementary Material for details). As a result, a total of 95 research articles were collected and screened for partition- and FS-based leakage (Supplementary Table S1, see also Supplementary Figure S1). Leakage related to DP and/or FS was identified in 41 and 34 of these, respectively (with 14 containing both), while potential leakage could not be reliably confirmed in a total of 13 and 10 cases given the provided description (i.e., unclear or unspecified in methods). Out of the total number of 119 evaluations, 25 and 74 contained no or with at least one form of leakage, respectively, not considering 20 cases with incomplete description. Collecting published classification accuracies, the median highest reported performance of 97.00% (IQR: [91.30, 98.99]) from leaky- in contrast to that of 86.00% (IQR: [75.77, 92.2]) from leakage-free pipelines supported our concern that performance metrics are likely overestimated due to leakage effects (p < 4.6098 × 10−7, z = 5.0419, Wilcoxon rank sum test, r = 0.5067). Notably, there were 6 papers (reporting 7 pipelines) that explicitly contrasted SZ detection performance in subject-based and epoch-based DP schemes (disregarding FS) on the same dataset [13, 15, 31–34]; synthesizing outcomes showed a statistically significant drop in performance (p = 9.5559 × 10−4, t6 = 6.0107, paired t-test) with an effect size of d = 2.9515 (CI:[1.4025, 4.4505]) when switching from epoch-based (99.04 ± 0.78%) to subject-based (88.06 ± 4.86%) evaluation (Supplementary Table S2). Finally, among the 95 screened studies, 28 analyzed private dataset, while 82 did so with one or more of three open EEG datasets ([35–37]).

Leaky vs leakage-free classification analyses

Accordingly, the observed trends warrant experimental validation. To this end, we implemented leaky and leakage-free realizations for various gold standard ML/DL pipelines employing both data-driven ‘black box’ and handcrafted EEG feature-based approaches. Performance of these models was evaluated on three openly available databases [35, 36, 38] (denoted MSU, RepOD and SU-SZ) containing resting-state EEG recordings from SZ patients and matched healthy control (HC) individuals (see Table 1 for dataset characteristics). Please see the Supplementary Material (ML-NICAD checklist) for details on hardware/software configurations used throughout this study.Table 1Basic demographic and EEG data collection characteristics of the analyzed datasets.

Dataset N Mean Age Sex EEG (#ch) sf (Hz) L (s)

MSU [35] HC 39 12.25 [11–13.75] N/A 16 128 60

SZ 45 12.25 [10.67–14] N/A

RepOD [36] HC 14 27.75 (3.16) 7F/7M 19 250 720

SZ 14 28.10 (3.72) 7F/7M

SU-SZ [38] HC 39 32.8 (9.5) 15F/24M 64 (19) 1000 (250) 100

SZ 38 34.3 (10.7) 17F/21M

For age, brackets and square brackets indicate standard deviation and range, respectively. For number of channels, values in brackets indicate the final number of channels included in the analysis. For sampling frequency, values in brackets indicate final sampling rate after downsampling. N/A indicates information not available.

HC healthy control, SZ schizophrenia, N sample size, #ch number of channels, sf sampling frequency, L maximum available data length in all subjects.

The effect of partition-based leakage on classification performance in data-driven pipeline using deep convolutional neural networks

In the first pipeline we utilized a pre-trained ResNet-18 [39] convolutional neural network (CNN) architecture fine-tuned via transfer learning [40]. Continuous EEG segments were spliced into 5 s epochs to augment sample size and epochs were transformed into channel-wise time-frequency features organized into a tri-color image-like structures reflecting scalp topology (see Methods). HC vs. SZ classification performance was evaluated in a stratified 10-fold cross-validation (CV) pipeline. Random DP was epoch-based (i.e., irrespective of subject ID) in the leaky implementation, while subject-based in the leakage-free scenario. SZ classification performances on the three datasets are summarized in Table 2. In the RepOD and SU-SZ databases accuracy dropped significantly from leaky to leakage-free scenarios (confirmed via permutation testing), while the performance drop was marginally significant (p = 0.0990) in the MSU dataset. The largest leaky vs. leakage free difference was observed in the RepOD dataset (28.27%), followed by the SU-SZ (12.79%) and MSU (4.03%) datasets.Table 2Classification performance reports from leaky and leakage-free implementations of the CNN-based models on the three analyzed datasets.

Dataset Leaky DP Leakage-free DP Difference (CI) p-value

MSU 85.32 ± 3.65% 81.30 ± 8.72% 4.03% (−0.69; 8.97) p = 0.0990

RepOD 86.33 ± 1.53% 58.06 ± 7.45% 28.27% (23.31; 33.57) p = 0.0001

SU-SZ 80.45 ± 4.76% 67.66 ± 10.95% 12.79% (6.51; 19.81) p = 0.0009

CNN convolutional neural network, DP data partition, CI confidence interval.

The effect of partition-based leakage on classification performance in ensemble-based pipeline using handcrafted features

While the data-driven approach confirmed the effect of partition-based leakage on classification performance, the purpose of the second pipeline was two-fold: i) to investigate the extent of leakage introduced by selecting features outside-the-loop and ii) to confirm that leakage is also prevalent even when features are manually engineered from EEG data instead statistically learned by the model. A boosted ensemble of decision trees was used as classifier, considered efficient in identifying complex interactions among diverse feature sets [41]. Continuous EEG was first segmented into 10 s epochs (to allow for better estimation of dynamic EEG properties) and five common feature types (frequency domain, time domain, non-linear, complexity and connectivity features, see Methods) were computed for each channel, resulting in a total of 386–458 features depending on the number of channels in the dataset. Model performance was evaluated in a 10-fold CV framework similarly as described above with DP-related leakage assessed as described previously. Additionally, we also employed the ReliefF algorithm [42] to select features that efficiently discriminate between HC and SZ. In leaky FS, top 50% features were first selected using the full dataset before 10-fold CV evaluation, while FS was performed within every CV iteration using only the training set in the leakage-free implementation. Since outcomes with these parameters indicated negligible effect of FS leakage in this pipeline, below we report the effect of leaky DP with leakage-free FS (Table 3), while more details are reported in the Supplementary Material and Supplementary Table S4.Table 3Classification performance reports from leaky and leakage-free implementations of the decision tree-based ensemble models on the three analyzed datasets.

Dataset Leaky DP Leakage-free DP Diff (CI) p-value

MSU 95.43 ± 4.31% 82.27 ± 7.91% 13.16% (6.36; 18.37) p = 0.0001

RepOD 97.47 ± 1.07% 64.73 ± 13.58% 32.74% (24.36; 42.66) p = 0.0002

SU-SZ 94.42 ± 3.24% 64.91 ± 8.96% 29.50% (24.30; 34.94) p = 0.0001

DP data partition, CI confidence interval.

Overestimation of classification performance was even more apparent than in the CNN-based pipeline. Namely, classification accuracy neared or surpassed 95% performance dropped significantly in all three datasets (average 25.13%) when switching to leakage-free implementations. Interestingly, while leaky performance was the highest on the RepOD dataset (97.47%), the most substantial drop in classification accuracy was also observed here (32.74%), indicating severe overfitting on subject-specific patterns.

Realistic subject-based evaluation: the effect of leaky feature-selection on classification performance

While these analyses clearly demonstrate inflation of SZ detection accuracy due to DP-based leakage, they do not yield realistic estimates on clinical performance, as model loss is minimized on epochs rather than aggregated subject-level assignments e.g., via majority vote [43]. Therefore, to mimic a realistic scenario, we re-used the previously obtained hand crafted feature set and characterized each subject with a single feature vector by averaging features over all epochs. Since this scenario eliminates the potential leakage from non-subject-based DP, here we only focused on addressing the effect of leaky FS. Furthermore, given the substantially smaller datasets, we employed a more radical FS pipeline, where among the features selected by ReliefF, only the top 10 were retained using the maximum relevance minimum redundancy (MRMR) technique [44]. Support vector machine (SVM) with radial basis function (RBF) kernel was used for classification. To make maximal use of available data, performance was evaluated in an exhaustive leave-one-subject-out (eLOSO) CV framework and leaky and leakage-free FS was implemented as described previously. As shown in Table 4, leaky FS resulted in significant overestimation of performance in all three datasets, with an average 8.72% improvement over leakage-free implementations.Table 4Classification performance reports from leaky and leakage-free implementations of the SVM-based models on the three analyzed datasets.

Dataset Leaky FS Leakage-free FS Diff (CI) p-value

MSU 90.94 ± 20.06% 80.77 ± 27.78% 10.17% (9.12; 11.23) p = 0.0001

RepOD 63.78 ± 33.04% 56.38 ± 35.77% 7.40% (2.30; 12.50) p = 0.0206

SU-SZ 81.44 ± 27.00% 72.84 ± 31.01% 8.60% (7.05; 10.17) p = 0.0001

SVM support vector machine, FS feature selection, CI confidence interval.

The cutting-edge approach: are transformer-based EEG foundation models immune to leakage-related performance overestimation?

Beyond conventional ML techniques, recent advances in artificial intelligence led to the emergence of so-called EEG foundation models, which use transformer-based architectures and arguably represent the current frontier of decoding EEG activity [45]. These models are trained using vast amounts (e.g., 10,000 + h) of diverse EEG data and appear to outperform former approaches on benchmark classification tasks; however, their performance under leaky vs. leakage-free scenarios has not been assessed yet, especially in the case of detecting SZ. For this purpose, we utilized the recently introduced REVE model [46], which provides state-of-the-art performance. Since EEG foundation models are data driven, we devised a similar evaluation scheme as in our CNN-based pipeline probing for DP-based leakage, using 5 s epochs as model input (see Methods). Embedding vectors produced by the REVE encoder were mean-aggregated to provide a single feature vector per EEG epoch, and an SVM with a linear kernel was used for classification. As presented in Table 5, leakage in DP resulted in significantly overestimated classification performance in case of all three datasets by an average 22.53%.Table 5Classification performance reports from leaky and leakage-free implementations of EEG foundation model-based approaches on the three analyzed datasets.

Dataset Leaky DP Leakage-free DP Difference (CI) p-value

MSU 96.43 ± 1.70% 83.47 ± 9.48% 12.96% (7.92; 18.31) p = 0.0003

RepOD 98.26 ± 0.31% 73.07 ± 20.00% 25.19% (13.74; 40.57) p = 0.0003

SU-SZ 88.80 ± 2.35% 59.36 ± 13.03% 29.44% (21.62; 36.47) p = 0.0001

DP data partition, CI confidence interval.

Performance comparison summary

As a summary, Fig. 2 illustrates model performances in all evaluation schemes through eight different performance metrics commonly employed in similar analyses (see Methods). Corresponding numerical values are provided in Supplementary Table S5 for completeness, while confusion matrices for the four evaluation schemes are illustrated on Supplementary Figures S2-S5. Model performances in each scenario are also illustrated via receiver operator characteristic curves on Supplementary Figure S6. It can be ubiquitously observed on Fig. 2 that for all four classification techniques that leaky evaluation pipelines (indicated via dashed lines) outperform leakage-free implementations (denoted via straight lines) in all aspects. In most cases, performance in leakage-free cases is balanced albeit poor, indicating low discriminative power. Notably, for the RepOD dataset all leakage-free techniques except the subject-based SVM have stronger sensitivity but low specificity and precision, indicating a bias towards the SZ class and high Type I. error.

While none of the four pipelines stand out as superior in leakage-free scenarios, it is worth noting that performances strongly depend on the dataset at hand. Specifically, while in leaky DP evaluations (i.e., not considering the subject-based SVM pipeline) performances are comparable among the three datasets, a great variability emerges when DP-related leakage is eliminated. The MSU dataset stands out as the most separable with balanced accuracy at or above 80% in all leakage-free evaluations. On the other hand, leaky performances surpassing 90% drop sharply to the 60–70% range (or even below 60%, depending on model architecture) for the RepOD and SU-SZ datasets. This performance range is somewhat lower than most outcomes reported in the literature where the risk of leakage is low (median of 86% accuracy from 25 studies). The most likely reason for this is that we deliberately did not employ hyperparameter tuning (HT), in part to gain insight on between-dataset generalization, but also as HT can be another source of information leakage (see Discussion). Therefore, while leakage-free evaluation performance in the 80–90% range appears achievable for single datasets, ‘realistic’ performance of current models is most likely below that due to poor between-dataset generalizability.

How partition-based leakage inflates performance: within- vs. between-subject variability in EEG

While the erroneous nature of utilizing the complete dataset for selecting most discriminative features is straightforward, the issue is less obvious with regards to partitioning. Specifically, epoch-based partitioning of data can cause leakage if unique, distinguishing patterns are present in the data that are characteristic of the individual instead of their group. On the level of data this can manifest as smaller within-subject compared to between-subject variability. We performed auxiliary analyses to confirm and illustrate this both for native EEG signals as well as extracted signal features (see Methods). Riemannian geometry-based analysis of covariance-transformed EEG epochs confirmed that within-subject spread of samples was significantly smaller (p < 10−5 in all cases) than between-subject (group-level or global) variability in the data for all three datasets (Fig. 3 upper panels, see also Methods). Additionally, 2-dimensional projection of multi-dimensional feature vectors via t-distributed stochastic neighborhood embedding (t-SNE) [47] indicated that same-subject samples form tight, non-overlapping clusters, while HC and SZ groups cannot be clearly delineated (Fig. 3 lower panels). Both analyses support that the dependence of EEG samples obtained from the same subject is stronger than group association, likely due to subject-specific characteristics. Therefore, there is real risk that an ML/DL model will learn to recognize these patterns and use them for decision making – instead of less potent group-characteristic patterns – if this dependency is not considered during DP.

Discussion

The two cases of information leakage considered in this study – namely, DP- and FS-related leakage – are both categorized as explicit cases of “test-to-train” leakage in the recent survey of Sasse and colleagues [18]. Specifically, ranking and selecting features using the complete dataset (i.e., outside-the-loop) directly exploits information from prospective test samples in the model training, leading to falsely inflated performance. This effect was addressed formerly by Shim and coworkers [20] on simulated data and EEG from HC and PTSD cohorts, showing that outside-the-loop FS overestimated classification accuracy on average by ~9% compared to inside-the-loop FS. This aligns well with our findings (8.72 ± 1.39%) in the case of SZ detection. Our analyses indicate that effects stemming from DP can be even more severe (15.03 ± 12.27 and 25.20 ± 10.38%). This is somewhat smaller than those reported in previous SZ literature [13, 15, 31–34] (10.98 ± 4.83% on average), although none of these studies attributed this effect to information leakage. On the other hand, Del Pup and colleagues [19] recently compared epoch- vs. subject-based pipelines to classify Parkinson’s and Alzheimer’s patients against HC and reported substantial drops in balanced accuracy (ranging ~15–45%). A fundamental theoretical assumption of supervised learning approaches is that different examples in the dataset are independently and identically distributed (I.I.D.) and sampled from a fixed probability distribution [18]. The I.I.D. assumption is violated by multiple EEG samples (epochs) from the same individual, as they likely contain the same characteristic traits specific to the individual. This was supported by our auxiliary analyses, even when attempting to match data distributions among subjects (Fig. 3). As a result, it is possible that an improperly trained model will recognize the individual (i.e., Jane Doe/John Smith) as their characteristic patterns were encountered during the training phase and simply predict the already associated group label instead of making this decision based on group-specific (i.e., HC/SZ) traits. Another important consequence of this phenomenon is that a model relying on subject-characteristic traits will likely categorize new, previously unseen subjects poorly, whose data were not encountered during the training procedure. In that regard, none of the reviewed articles reported cross-dataset classification performances. Overall, our findings emphasize an issue verifiably present in current literature, and while we can only conjecture, it is likely that performance metrics in the 42 other sampled studies employing only epoch-based DP are overestimated to a similar extent.

There is yet another aspect beyond DP and FS that we did not consider here, namely tuning hyperparameters of the classifiers. Training most DL models involve the heuristic selection of numerous hyperparameters (e.g., learning rate, optimization algorithm, hidden layer architecture, or regularization [48]) and optimal hyperparameter settings most often yield substantially better performances [49]. While HT can be performed both outside- and inside-the-loop similarly to FS, here we decided not to consider nor employ it for reasons outlined below. As emphasized by Del Pup and colleagues recently [19], training DL models for EEG classification should utilize a nested CV pipeline, where in each CV iteration (outer loop), an internal CV (inner loop) is performed by further partitioning the training set for HT. To facilitate an exact comparison between all our cases, we omitted this step in our CNN-based pipeline and instead model hyperparameters were either set to defaults or according to common ML/DL practices. This is likely the reason why our models – even with leaky implementations – underperformed compared to similar approaches in the literature (e.g., [50]), where HT was employed. It should be emphasized, however, that i) hyperparameter tuning, if not employed in a nested CV framework, is expected to have a similar effect on performance as leaky FS, and ii) excessive hyperparameter tuning on limited small datasets even with best practices can result in overfitting to dataset characteristics, resulting in poor generalization to new data [51]. Considering the excessive computational burden associated with HT, as well as to maintain clarity and conciseness of this paper, we eventually decided not to explore the effects of HT here (similar to previous works, see e.g., [25]). Nevertheless, we acknowledge this as a current limitation and an important research question that we intend to pursue in future work.

Further considerations should be made when working with openly available datasets. Recent work indicates that optimal parameter settings yielding the highest performance are often specific to the dataset at hand [49]. Therefore, it is probable that performance metrics obtained on these relatively small, open EEG databases (80 evaluations from the literature sample, as well as our analyses) would not hold when final models were tested on newly obtained data. Second, excessive use of open datasets invokes the issue of “dataset decay” [18]: when multiple hypotheses are tested on the same dataset, the chance of Type I error and thus false discoveries increases. As a reference, 43 out of 95 papers in our sample analyzed the RepOD dataset. Furthermore, given available reports on the same dataset, current/future research is incentivized to push for even higher performances; this can potentially result in models that overfit on dataset idiosyncrasies besides disease-specific traits, or employing features based on a priori knowledge of the same given dataset. Consequently, there is increased likelihood that some of the outstanding performances are results of chance or influenced by double dipping, also diminishing generalizability of new findings due to dataset over-usage [18]; however, we must note that this is not explicitly supported by data here. Finally, our analyses show substantial variability in leakage-free same-model performances for the three datasets, likely stemming from differences in the data collection environment and subject cohort characteristics. Importantly, this variability is much smaller when leakage is present, indicating that a single, outstanding performance report obtained from a singular dataset is no guarantee for true, clinically relevant generalizability. These notions all emphasize the need for open research efforts and especially data sharing. In this regard, the SU-SZ dataset, which was published in 2025 and at the time of writing this manuscript have not yet used in an EEG-based SZ detection study [38], might provide a new opportunity to test the generalizability of previously developed classification approaches.

We would also like to briefly address data reporting standards in relevant literature. According to our sampling, 22 out of 95 articles provided insufficient information in their methods description to confirm if e.g., if DP was epoch- or subject-based, or if FS was performed inside or outside the CV loop. This proportion is much higher when one aims for an exact replication of the analytical and evaluation pipeline e.g., the full parameter space scanned during HT is incompletely reported. Access to code implementations is rarely provided (only 3 out of 95 sampled cases), preventing verification of the outcomes.

To propel EEG-based CAD systems in SZ towards true clinical utility, there appears to be a need for establishing best reporting practices considering both dataset, analytical and evaluation pipeline aspects. To this end, below we summarize the most relevant takeaways substantiated by our analyses in Table 6. Furthermore, to facilitate consistency among independent studies and transparency in reporting methods and results, we propose a checklist template, which we termed the Checklist for Transparent Reporting of Machine Learning Pipelines in Neuroimaging-based Computer-Aided Diagnostics (ML-NICAD Checklist). This checklist is composed of two sections. The first allows for a concise summary of the article where the outcomes are reported, including core details of the analyzed dataset(s). The second concerns common building blocks of classification pipelines, including i) model specifics and hardware/software environment, ii) data preprocessing, iii) feature extraction/engineering, iv) model architecture and training, v) hyperparameter tuning vi) data partitioning and cross-validation, vii) feature selection, and viii) performance evaluation. The blank document can be accessed as Supplementary Document 1 of this submission. It is important to acknowledge that not all classification approaches might fit the same rigid framework and thus we intend the ML-NICAD checklist to be customizable; as an illustration, we provide the completed checklists detailing all evaluation pipelines presented in this study in Supplementary Document 2.Table 6Best practices and recommendations for developing EEG-based CAD systems.

Aspect Recommendation

Data Partitioning In case of data augmentation via EEG segmentation, data should always be partitioned along subjects instead of treating epochs as independent samples.

Feature Selection Feature selection should always be performed excluding test data and using only the training set (inside-the-loop for CV).

Hyperparameter tuning Hyperparameter tuning should always be performed along the same lines as FS. Since HT also requires validation data, employ nested, subject-based CV as recommended by Del Pup and colleagues [19].

Generalization Verify classification pipelines on multiple, independent datasets to better assess the generalizability of the proposed approach across data collection environments and patient cohorts.

Transparency Provide a detailed, structured summary of the evaluation scheme, with particular emphasis on how DP, FS and HT were handled (e.g., via the ML-NICAD checklist), if applicable.

Reproducibility Whenever possible, share analysis codes for validation and reproducibility.

Dataset Decay Whenever possible, share anonymized EEG data to facilitate development and cross-validation of EEG-based CAD systems for SZ detection.

Lastly, we address the limitations and broader implications of this present study. Foremost, the central aim for this work was not to propose a benchmark model for EEG-based SZ detection, but to emphasize the potential effects of leaky evaluation protocols. As such, the reported leakage-free performances should not be treated as ‘state-of-the-art’, especially that certain steps (most particularly, HT) were omitted. We only focused on leakage stemming from DP and FS as we believed these being to most pertinent in literature; there are, however, numerous other sources of leakage to attend to, as summarized in [18]. Performances in the CNN-based pipeline could be further improved via nested CV scheme involving HT [19]. Here we only analyzed previously shared, confined EEG datasets; none of the models were evaluated on external samples, and cross-dataset or pooled-dataset performances were not explored. Small sample sizes also limit the generalizability of our findings. We only evaluated four gold standard ML/DL approaches here, while other techniques (e.g., recurrent neural networks [52] or other transformer-based architectures [53] besides EEG foundation models) also show promise. Note, however, that the leakage mechanisms considered here (also including HT) are model agnostic; therefore, they are expected to have similar effects in a broad range of approaches where these notions are applicable. Evaluations only considered resting-state EEG databases for the sake of comparability, although task-specific EEG patterns can provide additional opportunities for SZ detection such as exploiting evoked response differences [37]. We also did not account for demographic differences such as the MSU dataset [35] consisting of adolescents instead of adults, or medication effects such as in the SU-SZ dataset. Our literature screening was based on recent systematic reviews [6, 7, 27, 28], discarding works that were published outside the inclusion period of these papers. Therefore, we likely omit a substantial amount of recent research, given the increasing trend in the number of related publications (see also Supplementary Figure S1) and considered time frame. As an exploration, we performed a screening of literature published after January 1st, 2024, using the Rayyan [54] tool (see Supplementary Material). After removing duplicates, screening of 219 titles and abstracts indicated 125 articles for full-text consideration. While performing an updated systematic or scoping review to synthesize these works is beyond our current scope, these trends clearly indicate the heightened interest in EEG-based SZ detection and warrant frequent and structured synthesis of relevant literature. Finally, we should like to emphasize that while we investigated the issue of leakage in the case of SZ detection, the implications of our findings are not limited to this particular problem, and similar concerns are likely to arise across a wide range of EEG-based psychiatric or neurological applications, as already indicated by the literature [22, 26]. Furthermore, these issues are not unique to EEG as a modality, and can emerge whenever limited sample sizes are countered with segmentation-based data augmentation or excessive feature engineering e.g., in fMRI-based studies [25].

Methods

Literature review

As there are multiple recent systematic reviews on EEG-based detection of SZ [6, 7, 27, 28], we synthesized all the articles considered by these works instead of performing another systematic review. While the work of Khare and colleagues [30] is not a systematic review, it contains a survey of over 40 papers (see Table 1 in [30]) and thus we decided to also include it in the synthesis. After eliminating duplicates and inadmissible works (see Supplementary Material, section Notes for details), this procedure yielded 95 original publications (Supplementary Table S1), including a total of 119 evaluation pipelines (some works analyzed multiple datasets). The full text from all these publications was assessed. We followed the screening guidelines and categorization recently proposed by Young and colleagues [24] to assess if leakage was i) present, ii) absent, or iii) it could not be determined with reasonable certainty. Please note that these categories are denoted as high, low or moderate risk for leakage, respectively, in the original work [24]. Classification accuracy (or in cases of multiple evaluations, the highest reported performance) was extracted from each evaluation pipeline as the characteristic performance metric. For more details on the assessment of leakage risk, please see the Supplementary Material.

Experimental datasets

We analyzed recordings from three, open EEG repositories, detailed below. In all datasets, individuals in the SZ cohorts had a diagnosis of SZ according to international ICD-10 (F-20, F-21 or F-25) established by a psychiatrist. We also report on medication information where available.

Dataset 1: MSU dataset

The first dataset was collected at the M.V. Lomonosov Moscow State University in Moscow, Russia, and shared by Borisov and colleagues [35, 55], containing one minute of eyes-closed resting-state EEG data from n = 39 HC (mean: 12.25 years, range: 11–13.75 years) and n = 45 SZ (mean: 12.25 years, range: 10.67–14 years) adolescents. EEG was collected at 128 Hz sampling from 16 locations (standard 10–20 locations excluding Fp1, Fp2 and Fz) referenced to linked earlobes. The original study was approved by the ethical committee of Moscow Mental Health Scientific Center of Russian Academy of medical Sciences, and written informed consent was collected from parents of the participants [56]. No information was available on the distribution of females/males in the study groups, or if SZ participants were on medication at the time of recording. The dataset is available at http://brain.bio.msu.ru/eeg_schizophrenia.htm.

Dataset 2: RepOD dataset

The second dataset was collected at the Institute of Psychiatry and Neurology in Warsaw, Poland, and shared by Olejarczyk and Jernajczyk [36], containing ~15 min of eyes-closed resting-state data from n = 14 HC (7 females aged 28.7 ± 3.4 years and 7 males aged 26.8 ± 2.9 years) individuals and n = 14 patients with paranoid SZ (7 females aged 28.3 ± 4.1 years and 7 males aged 27.9 ± 3.3 years). Data was sampled at 250 Hz from 19 standard locations of the international 10–20 montage [57], with the reference electrode positioned over FCz. The Ethics Committee of the Institute of Psychiatry and Neurology in Warsaw reviewed and approved the study protocol and all participants provided written informed consent before recordings. Individuals in the SZ cohort were subjected to a medication washout period with a minimum of seven days before EEG collection. The dataset can be accessed at https://doi.org/10.18150/repod.0107441.

Dataset 3: SU-SZ dataset

The third dataset was collected at the Department of Psychiatry and Psychotherapy, Semmelweis University in Budapest, Hungary, and shared by Racz and colleagues [38]. The full dataset contained 2 min of eyes-closed resting-state data from n = 39 HC individuals (age: 32.8 ± 9.5 years, 15 females) and n = 38 patients with SZ (age: 34.3 ± 10.7 years, 17 females). Original data was collected from 64 scalp locations at 1000 Hz sampling; however, for dataset comparability, only the subset of 19 channels according to the 10–20 montage was considered in this study and the data was downsampled to 250 Hz (see below). The Semmelweis University Regional and Institutional Committee of Science and Research Ethics reviewed and approved the study protocol, and all participants provided written informed consent before enrollment. Participants were ON medication at the time of EEG collection. The dataset is available at https://doi.org/10.5281/zenodo.14808295.

Ethics statement

No original data from human participants was recorded for this manuscript; all analyzed data was obtained from repositories available in the public domain. All original studies were conducted in line with the Declaration of Helsinki, approved by their respective regional ethics committees (see details above), and written informed consent was obtained from all participants. The current study was approved by the Semmelweis University Regional and Institutional Committee of Science and Research Ethics (registration number: 197/2015) and was conducted in line with the Declaration of Helsinki.

EEG preprocessing

Pre-processing of raw EEG data was carried out in MATLAB 2025b (The MathWorks, Natick, MA) using the EEGLAB toolbox [58] version 2024.0. The same pre-processing steps were applied in all three datasets, with exceptions noted when applicable. From the MSU dataset, all available 1 min EEG was considered, while from the RepOD and SU-SZ datasets the first 720 s and 100 s of EEG recordings were selected, respectively, being the longest available segment for all participants in the corresponding datasets. Raw, continuous EEG data was first band-pass filtered between 1–45 Hz using a finite impulse-response filter using default EEGLAB settings. Additionally, the CleanLine algorithm [59] was used to eliminate line noise at 50 and 100 Hz. The 1000 Hz data of the SU-SZ dataset was downsampled to 250 Hz for computational efficiency and comparability with the other datasets. Then, independent component analysis (ICA) [60] was utilized to remove artifact signal components such as those related to eye movements, muscle or cardiac activity. The multiple artifact rejection algorithm [61] was used to identify artifact-related independent components based on six temporal, spatial and spectral features, which were removed before regaining the signal via reverse-ICA. Finally, data was re-referenced to the common average electrode, and signal for each channel was standardized to have zero mean and unit variance, thus eliminating spatial variability due to electrode impedances.

Deep learning pipeline using image-transformed EEG features

The advantage of DL classification is that it does not require extensive feature engineering; instead, complex models, such as CNNs, learn features from data directly. In particular, CNNs are most commonly used in computer vision applications [62], as early convolutional filters are effective in learning simple spatial patterns (such as edges and primitive shapes), while those in later layers of the neural network synthesize them to encode more complex spatial structures. In general, CNNs are effective on any type of data where information can be encoded as spatial or spatiotemporal patterns, and they are widely used in EEG classification based on time-frequency features [63]. On the downside, since CNNs usually involve training of millions of model parameters, they require substantial amounts of training data. To circumvent this, data augmentation techniques are commonly employed, such as segmenting continuous EEG into shorter epochs and thus multiplying sample size [63]. Additionally, when training data is scarce, transfer learning [40] is often employed, only retraining the final, fully connected layer(s) of a deep CNN already trained on millions of samples, thus preserving the model’s capability to extract and recognize low-level features. Accordingly, in this pipeline we first segmented continuous EEG into short epochs, transformed EEG them into ‘image-like’ time-frequency features encoding spatiotemporal information, and fine-tuned a pre-trained CNN architecture for EEG epoch classification. Our analytical pipeline is a modified version of that proposed by Khare and colleagues [50].

EEG transformation

Continuous, pre-processed EEG was segmented into 5 s, non-overlapping epochs. From these, time-frequency information was extracted for every channel in three separate ways, much like in [50]: spectrogram using the Welch periodogram method, scalogram using continuous wavelet transformation (CWT), and smoothed pseudo Wigner-Ville Distribution (WVD). The spectrogram was estimated using the spectrogram() MATLAB function with a window size of 1 s, overlap of 90% (0.1 s step size), Hanning-windowing and a frequency resolution of 0.25 Hz. Complex-valued spectrograms S were converted into dB according to \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${dB}=20* {\log }_{10}(\left|S\right|)$$\end{document}dB=20*log10(S). CWT transformation was performed using the cwt() MATLAB function with default Morse wavelet, and WVD was performed using the wvd() MATLAB function with “smoothedPseudo” option. Both CWT and WVD outputs were log-transformed for better scaling. From all three time-frequency decompositions, data from outside the 1–45 Hz regime were excluded. The obtained channel-wise spectrogram, CWT and WVD matrices were then arranged into a tri-color image the following way. For the MSU dataset, a 4 × 5 grid was created, while a 5 × 5 for the RepOD and SU-SZ datasets, due to the slightly different channel arrangement (16 vs. 19 channels). For all three methods, time-frequency matrices obtained from each channel were assigned to grid positions corresponding to their approximate arrangement on the scalp, as illustrated on Fig. 4A, with grid cells not corresponding to an existing scalp location filled using average time-frequency matrices of nearby channels. These composite images were then rescaled to 224 × 224 using the imresize() MATLAB function and scaled between [0, 255] to fit input requirements for the ResNet-18 architecture (see below). Finally, these processed matrices from spectrogram, CWT and WVD were organized into a tri-color image of size 224 × 224 × 3 (Fig. 4B), capturing short-term spatiotemporal dynamics in EEG. The procedure was carried out for all non-overlapping epochs in all subjects, resulting in 12 samples per subject (1008 total) for the MSU dataset, while 144 (4032 total) samples and 20 (1540 total) samples for the RepOD and SU-SZ datasets, respectively.

Evaluation pipeline

We employed the widely utilized ResNet-18 [39] architecture, pre-trained on over one million images of the ImageNet dataset [64], as implemented in the MATLAB Deep Learning Toolbox imagePretrainedNetwork library. Since ResNet-18 has about 11.7 million parameters, we employed a transfer learning framework to prevent overfitting. In that, we froze weights in the first ten convolutional layers of the model, added a DropOut layer with probability 0.4 before the final fully connected layer, and replaced the classification layer to predict two outputs after applying SoftMax on the preceding fully connected layer outputs. The resulting architecture was then evaluated in two pipelines on all three datasets independently, using a k-fold CV framework, where in each of k folds, 1/k proportion of the data was set aside for evaluation and the model is trained on the remainder (1-1/k) of the data. In the leaky implementation, we employed a stratified k-fold CV evaluation in which data epochs were assigned to train and evaluation sets regardless of their subject ID but preserving HC-SZ proportions in both sets. In the leakage-free implementation, during each CV iteration the subjects were first partitioned into training and evaluation subjects, then all samples from a given subject were sorted exclusively into the training or evaluation sets. The MSU and SU-SZ datasets allowed for an approximate 10-fold CV framework in the leakage-free, subject-dependent implementation (e.g., 3–4 HC and SZ in the evaluation set) while preserving the approximate proportion of HC and SZ in each evaluation folds, however for the much smaller RepOD dataset we utilized 7-fold CV instead, yielding completely balanced partitions. In both leaky and leakage-free CV schemes, all samples of the dataset were utilized as an evaluation sample at least once, and exactly once.

Machine learning pipeline using handcrafted EEG features

Small datasets might prevent the full utilization of DL capabilities in learning features from data without risking overfitting. Instead, one might employ certain feature engineering techniques to extract relevant information from data guided by a priori knowledge [65]. This approach is also commonly utilized in EEG classification studies [66–69], complementing the previously described data-driven approach. In most cases, a wide range of features is obtained initially along different principles, in order to capture various aspects of complex EEG signals.

Feature extraction

In our approach, we computed five different types of EEG features commonly utilized in the SZ literature (see e.g., [70]): i) power spectral features, ii) time-frequency features, iii) time-domain features, iv) non-linear dynamic features and v) connectivity features. Pre-processed EEG was first segmented into non-overlapping epochs of 10 s to allow for a better characterization of dynamic properties. Then, various EEG features for each channel were computed as follows (for details and formulas, see Supplementary Material):i)-ii) power spectral and time-frequency features: in every epoch, for every channel a spectrogram was obtained using 1 s window size, Hanning-windowing, 75% overlap (0.25 s step size) and frequency resolution of 0.25 Hz. Spectral power was integrated over five canonical frequency bands (delta: 2–4 Hz, theta: 4–7 Hz, alpha: 7–12 Hz, beta: 12–30 Hz and gamma: 30–45 Hz) to obtain time-resolved band-limited power (BLP) estimates. Then, each frequency band was characterized by its mean and variance of BLP over time. This procedure yielded 10 features per channel, per epoch.

iii) time domain features: temporal complexity of broadband (1–45 Hz) EEG in each 10 s epoch was characterized via Hjorth parameters [71] activity, mobility and complexity, and the detrended fluctuation analysis (DFA) scaling exponent [72]. Since Hjorth activity is the variance of native EEG, this feature was computed from pre-processed EEG before standardization, while mobility and complexity were obtained from standardized EEG data. DFA scaling exponents were obtained using the efficient multi-channel method proposed in [73]. These procedures yielded 4 features per channel, per epoch.

iv) non-linear dynamic features: non-linearity was characterized in the five canonical frequency regimes (as above) using Permutation Entropy (PE) [74]. For this analysis, pre-processed continuous EEG data was first band-pass filtered to the corresponding frequency regime using a 4th order zero-phase Butterworth filter. Then, normalized PE was computed using the full 10 s EEG of each epoch, for every channel. Parameters of PE were set to match the frequency regime characteristics as outlined in [75] (Supplementary Table S3). This analysis produced 5 features per channel, per epoch.

v) connectivity features: we utilized the Phase-Lag Index (PLI) as our connectivity estimator [76]. In that, weighted adjacency matrices were constructed for each frequency band by computing pairwise PLI between all channel pairs. Following the approach in [67], matrices were density-thresholded [77] using a set of threshold values ranging from 15 to 50% to eliminate spurious connections, and matrices were collapsed over various threshold values to yield robust estimates. Finally, in every frequency band the node degree was estimated for each EEG channel, yielding 5 node degree estimates for each channel, per epoch.

The above-described procedure produced feature vectors \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${F}_{{ch}}^{t}\in {{\mathbb{R}}}^{1\times 24}$$\end{document}Fcht∈R1×24 for each channel ch and epoch t. Feature vectors were concatenated according to all channels, leading to training samples of size \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${F}^{t}\in {{\mathbb{R}}}^{1\times 384}$$\end{document}Ft∈R1×384 for the MSU dataset and \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${F}^{t}\in {{\mathbb{R}}}^{1\times 456}$$\end{document}Ft∈R1×456 for the RepOD and SU-SZ datasets. Every subject in the MSU dataset provided 6 samples (504 total), while 72 samples (2016 total) and 10 samples (770 total) for the RepOD and SU-SZ datasets.

Feature selection

The feature extraction procedure yielded 384 initial features per epoch in the MSU dataset, while 456 initial features in the RepOD and SU-SZ datasets. To reduce dimensionality and eliminate irrelevant features from our initial set, we performed feature selection using the ReliefF algorithm [42]. ReliefF is a supervised feature selection algorithm that iteratively samples and evaluates feature relevance based on the feature’s ability to differentiate nearest neighbors from the same class (hits) and different classes (misses). ReliefF assigns positive weights to features that consistently show small differences in values among hits but large differences among misses, while negative weights arise when differences are larger among hits than misses. Therefore, a negative weight indicates that a feature is irrelevant for the classification problem. Here we utilized ReliefF in a simple manner with the number of nearest neighbors set uniformly to 10. As we have found that ReliefF assigned positive weights to over 50% of initial features in every case, we decided to uniformly retain the top 50% of features for classification in every evaluation scenario and CV iteration. This resulted in a final feature set of 192 for the MSU, while 228 for the RepOD and SU-SZ datasets. Feature selection was implemented in leaky and leakage-free ways. In the leaky scenario, first all features were standardized across the complete dataset. Then, ReliefF-based feature selection was performed on the full dataset, and CV evaluation was executed using this reduced, fixed feature set. In the leakage-free pipeline, feature selection (including standardization of features) was performed in every CV iteration separately using only data from the training set, and this subset was then utilized to classify examples in the evaluation set.

Evaluation pipeline

In this pipeline we utilized an ensemble classifier of decision trees enhanced with “GentleBoost” for classifying examples obtained by the above-described procedure. Specifically, in every CV iteration an ensemble classifier of simple decision trees (with maximum number of splits set to 5) was trained using the fitensemble() method in MATLAB, with setting the learning method to GentleBoost, the number of learning cycles to 100, and the learning rate to 0.1. Similarly to the data-driven pipeline, k-fold CV was implemented in leaky and leakage-free (i.e., subject-based) manners. Therefore, this pipeline was ultimately executed in four different ways: i) Leaky DP and leaky feature selection, ii) leaky DP but leakage-free feature selection, iii) leakage-free DP but leaky feature selection, and iv) completely leakage-free. In the main text we report results from cases ii) and iv) for comparability with the CNN-based pipeline, while details from all analysis scenarios are provided in the Supplementary Material.

Single feature vector-based pipeline

This pipeline utilized the same feature set as obtained previously; however, for each subject to be characterized by a single feature vector, the obtained estimates were averaged over all epochs per subject. Accordingly, for this analysis the MSU dataset consisted of 84 examples (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${F}^{t}\in {{\mathbb{R}}}^{1\times 384}$$\end{document}Ft∈R1×384), the RepOD dataset of 28 examples (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${F}^{t}\in {{\mathbb{R}}}^{1\times 456}$$\end{document}Ft∈R1×456), and the SU-SZ dataset of 77 examples (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${F}^{t}\in {{\mathbb{R}}}^{1\times 456}$$\end{document}Ft∈R1×456). Given the disproportionally large number of initial features compared to the drastically reduced sample size, we employed a more stringent feature selection and CV framework. However, since each subject contributed only one sample, DP-based leakage was not considered in this framework.

Feature selection

As an initial step, features were standardized. Then, the ReliefF algorithm was utilized to eliminate features with negative weight and thus likely irrelevant for classification. Finally, the number of features were further reduced by employing the maximum relevance minimum redundancy (mRMR) algorithm [44]. The mRMR is an iterative algorithm that aims to select a subset of features that are maximally relevant to the target variable to be predicted, while at the same time minimally redundant amongst themselves. Here, we set mRMR to identify the top 10 most relevant features in all cases, which were then used for training and classification.

Evaluation pipeline

Given the small samples, in this pipeline a support vector machine with radial basis function kernel was utilized for classification via the fitcsvm() MATLAB function. To maximally utilize the available datasets, we employed an extensive leave-one-subject-out (eLOSO) CV scheme. In that, the evaluation set was always defined as a single pair of HC-SZ individuals with all remaining data used for training, and all possible pair combinations – i.e., NHC×NSZ where NHC and NSZ are the number of HC and SZ individuals in the dataset, respectively – were utilized as evaluation cases. Consequently, this pipeline only considered leakage related to feature selection. In the leaky implementation, standardization and ReliefF+mRMR was performed first (i.e., outside-the-loop) to obtain the top 10 most relevant features, and eLOSO-CV was executed using this fixed feature subset in every iteration. In the leakage-free implementation, all feature selection steps – including standardization – was performed within every CV iteration (i.e., inside-the-loop) using only training set data, and the obtained feature subset was then selected from the evaluation set. Thus, the top 10 features varied for the different evaluation HC-SZ pairs in the CV evaluation. In both cases, classifier performance was characterized via classification accuracy averaged over CV iterations.

EEG foundation model pipeline

EEG foundation models (FMs) employ a similar label-free, self-supervised learning strategy as modern large language and multimodal models [45] in an attempt to identify meaningful patterns and salient features in EEG data. The end-to-end training procedure of these models (often referred to as pretraining) generally involves a transformer-based autoencoder architecture: an encoder projects native, unlabeled EEG data into a latent, high-dimensional embedding space, while a decoder attempts to reconstruct masked EEG segments (or their specific properties e.g., spectral characteristics) from these embeddings. Given sufficient amount of diverse pretraining data, a product of this procedure is an encoder that compresses EEG into meaningful embeddings, which meaningfully capture the intrinsic properties of brain activity. These embeddings then can be used in various supervised classification tasks (often termed downstream tasks) such as brain-computer interface applications or clinical decision making support, and thus FMs can be considered as “off-the-shelf feature extractors” [45]. While the pretraining is expensive and relies on thousands of hours of unlabeled EEG data, the downstream classifiers are usually simple and lightweight compared to the encoder itself. Overall, state-of-the-art FMs are built to recognize and extract inherent properties of EEG collected from a diverse set of conditions and experimental setups, to generalize well to unique or custom electrode montages, and to circumvent the bottleneck of manual data annotation, and therefore they hold great promise for future research dealing with EEG-based applications [45].

Conceptually, the embedding procedure performed by the pretrained encoder is feature extraction itself, and therefore downstream tasks do not commonly involve explicit FS. However, since the downstream classifier is ultimately trained to separate embedding vectors, it can be subject to DP-related effects same as any other classification approach. To the best of our knowledge, no previous study has investigated this issue yet on downstream classification tasks. For this purpose, we opted to utilize REVE [46], a recently proposed FM pretrained on over 60,000 h of both clinical and other EEG data. Note that according to the current documentation of REVE, none of the three SZ-EEG datasets considered in our study were part of the REVE pretraining dataset (see Appendix B in [46]).

EEG preparation

Preprocessed EEG was further prepared to align with data used for pre-training REVE [46]. In that, EEG was first resampled to have 200 Hz sampling rate (upsampling for MSU, downsampling for RepOD and SU-SZ) using the resample() MATLAB function, and continuous recordings for each channel were standardized to have zero mean and unit variance (subject-level). Then, recordings were segmented into 5 second-long, non-overlapping epochs. Finally, epochs with values outside the ±15 standard deviation regime were excluded. This resulted in a final sample size of 1008, 3909 and 1500 EEG epochs for the MSU, RepOD and SU-SZ datasets, respectively, which were subsequently used as inputs to the REVE encoder.

Feature extraction

We used the REVE base model (accessible at https://huggingface.co/brain-bzh/reve-base) to transform the 5-second EEG epochs into 512-dimensional embedding vectors. Parameters of the REVE model were kept frozen during the procedure. Since REVE internally segments EEG into 1-second non-overlapping vectors, this procedure yielded 5 embedding vectors per EEG epoch, which were then aggregated into a single feature vector of size 1 × 512 via mean-pooling. Accordingly, the final samples for the MSU, RepOD and SU-SZ datasets were of size 1008 × 512, 3909 × 512 and 1500 × 512, respectively.

Evaluation pipeline

To remain consistent with standard practices in FM performance evaluation [45], we performed linear probing by using SVMs with linear kernels for classification. For this, we used the fitcsvm() MATLAB function with ‘KernelFunction‘ set to ‘linear‘ and ‘Standardize‘ set to true. The full, 512-dimensional embedding vectors were used as model input and no FS was performed. Similar to the CNN-based evaluation pipeline, stratified 10-fold CV was utilized for the MSU and SU-SZ datasets, while 7-fold CV for the RepOD dataset in both leaky DP (random split) or leakage-free (subject-based split) implementations.

Performance evaluation

In the primary analyses, model performances in every CV iteration were characterized via classification accuracy i.e., the proportion of correctly predicted evaluation samples to the total number of evaluation samples. For a more comprehensive assessment, ground truth labels and predictions were pooled from all CV iterations and performance was characterized through eight standard metrics [78]. With TP, TN, FP and FN denoting true positives, true negatives, false positives and false negatives, respectively, the following measures were computed:i.\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${\rm{Balanced\; Accuracy}}( \% ):\frac{{Sensitivity}+{Specificity}}{2}\times 100$$\end{document}Balanced Accuracy(%):Sensitivity+Specificity2×100

ii.\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${\rm{Specificity}}( \% ):\frac{{TN}}{{TN}+{FP}}\times 100$$\end{document}Specificity(%):TNTN+FP×100

iii.\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${\rm{Sensitivity}}( \% ):\frac{{TP}}{{TP}+{FN}}\times 100$$\end{document}Sensitivity(%):TPTP+FN×100

iv.\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${\rm{Precision\; or\; Positive\; Predictive\; Value}}( \% ):\frac{{TP}}{{TP}+{FP}}\times 100$$\end{document}Precision or Positive Predictive Value(%):TPTP+FP×100

v.\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${\rm{F}}1{\rm{Score}}( \% ):\frac{{Precision}\times {Sensitivity}}{{Precision}+{Sensitivity}}\times 100$$\end{document}F1 Score(%):Precision×SensitivityPrecision+Sensitivity×100

vi.\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$${\rm{Matthew}}{\rm{\mbox{'}}}{\rm{s\; Correlation\; Coefficient}}( \% ):\frac{{TP}\times {TN}-{FP}\times {FN}}{\sqrt{({TP}+{FP})({TP}+{FN})({TN}+{FP})({TN}+{FN})}}\times 100$$\end{document}Matthew’s Correlation Coefficient(%):TP×TN−FP×FN(TP+FP)(TP+FN)(TN+FP)(TN+FN)×100

vii.Cohen’s Kappa [79]

viii.receiver operator characteristic area under the curve (ROC-AUC) [80].

Riemannian geometry and assessment of data variability

The toolset of Riemannian geometry provides a suitable way to address variability in multi-channel EEG data [81]. In the Riemannian framework, multivariate data is first transformed into covariance matrix representation (being symmetric positive definite) and then the distance between different matrices can be captured via the Riemannian distance (or geodesic) that is invariant under affine transformations [82]. We analyzed the EEG epochs utilized in the CNN-based evaluation, and each epoch was transformed into trace-normalized covariance matrix structure using the Ledoit-Wolf method [83]. In each dataset, individual covariance matrices were re-centered using a reference matrix obtained as the Riemannian (Karcher) mean of all covariance matrices in the dataset [84]. This step is equivalent to standardization for univariate signals, and its goal was to match distributions among different subjects and thus minimize between-subject variability. Then, the following prototype (i.e., center of mass) matrices were computed from corresponding epochs as in [84]: i) subject centroids, obtained as the Riemannian mean of covariance matrices from that subject, ii) group centroids, obtained as the Riemannian mean of all covariance matrices from HC/SZ groups, and iii) global centroids, obtained as the Riemannian mean of all covariances in the dataset, irrespective of group label. Within-subject variability was defined as the average Riemannian distance of a subject’s samples to their own centroid, while between-subject and global variability for each subject was defined as the average Riemannian distance of the subjects’ samples to their respective group centroid and the global dataset centroid, respectively. Note that this approach is analogous to within- and between-class variability assessment in traditional clustering analysis, however the Riemannian approach allows for exploiting multivariate interactions in the data via the covariance representation and thus is more suitable to address variability in multi-channel EEG data. For more details on the application of Riemannian geometry on EEG data and related concepts, please see [81] or [85].

For variability in extracted EEG features, the t-SNE [47] method was utilized, which is a common visualization and analysis technique used in data science. In essence, t-SNE employs a non-linear approach to map high-dimensional data to a lower-dimensional (e.g., 2D) space while preserving local similarities, and thus it is ideal in intuitive exploration of internal data structures such as clustering of samples or separability of classes [47]. We analyzed the EEG feature vectors derived for the ensemble-based evaluation for this analysis. All features were standardized using all data in the given dataset to reduce distribution shifts across subjects and ensure similar scaling of all features. The tsne() MATLAB function was utilized with default perplexity value of 30 to obtain two-dimensional embeddings of full feature vectors.

Note that in both analyses, data processing steps (re-centering, standardization) were employed to reduce between-subject distribution shifts in the data and therefore the resulting outcomes should be considered as conservative estimates on data variability.

Statistical testing

The performance drops from leaky to leakage-free implementations were characterized as \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$\mu \left({{ACC}}_{L}\right)-\mu ({{ACC}}_{{LF}})$$\end{document}μACCL−μ(ACCLF) where μ(∙) denotes the average taken over CV iterations, and ACCL and ACCLF refers to classification accuracy in the leaky and leakage-free implementation, respectively. Confidence intervals for this difference were obtained via bootstrapping. In that, for both the leaky and leakage-free fold accuracies, an equal number of samples to the number of fold iterations (i.e., 10 in 10-fold CV, NHC×NSZ in eLOSO-CV) were chosen randomly with replacement, and the difference between average bootstrap accuracies were computed. This procedure was repeated K = 10,000 times. Then, the 2.5 and 97.5% percentiles of the obtained distribution were chosen as confidence interval boundaries. The statistical significance of the difference was estimated via permutation testing. In that, for each evaluation scheme, the leaky and leakage-free fold accuracies were first concatenated, then the complete set was shuffled before re-partitioning into leaky and leakage-free outcomes, with finally the difference between average leaky and leakage-free CV accuracies computed from these permuted sets. The procedure was repeated for K = 10,000 times, and the non-parametric p-value was defined as the proportion of permuted differences larger than the observed true difference. In the analysis of data variability, within-subject, between-subject and global variability values were contrasted using paired t tests or Wilcoxon signed rank tests, depending on data normality. Outcomes were adjusted for multiple comparisons (n = 3) using Bonferroni’s method at α = 0.05.

Supplementary information

References

1.Solmi M Seitidis G Mavridis D Correll CU Dragioti E Guimond S Incidence, prevalence, and global burden of schizophrenia - data, with critical appraisal, from the global burden of disease (GBD) 2019 Mol Psychiatry 2023 28 5319 27 10.1038/s41380-023-02138-4 37500825 Solmi M, Seitidis G, Mavridis D, Correll CU, Dragioti E, Guimond S, et al. Incidence, prevalence, and global burden of schizophrenia - data, with critical appraisal, from the global burden of disease (GBD) 2019. Mol Psychiatry. 2023;28:5319–27.10.1038/s41380-023-02138-4 37500825

2.Misiak B Samochowiec J Kowalski K Gaebel W Bassetti CL Chan A The future of diagnosis in clinical neurosciences: comparing multiple sclerosis and schizophrenia Eur Psychiatry 2023 66 e58 37476977 10.1192/j.eurpsy.2023.2432 PMC10486256 Misiak B, Samochowiec J, Kowalski K, Gaebel W, Bassetti CL, Chan A, et al. The future of diagnosis in clinical neurosciences: comparing multiple sclerosis and schizophrenia. Eur Psychiatry. 2023;66:e58.10.1192/j.eurpsy.2023.2432 PMC10486256 37476977

3.Oso TA Okesanya OJ Adebayo UO Obadeyi KB Musa SS Ahmed MM Advancing our understanding of schizophrenia: insights from recent research, emerging therapies, and future directions Exploration Neurosci 2025 4 1006112 10.37349/en.2025.1006112 Oso TA, Okesanya OJ, Adebayo UO, Obadeyi KB, Musa SS, Ahmed MM, et al. Advancing our understanding of schizophrenia: insights from recent research, emerging therapies, and future directions. Exploration Neurosci. 2025;4:1006112.

4.Etkin A Mathalon DH Bringing imaging biomarkers into clinical reality in psychiatry JAMA Psychiatry 2024 81 1142 7 39230917 10.1001/jamapsychiatry.2024.2553 Etkin A, Mathalon DH. Bringing imaging biomarkers into clinical reality in psychiatry. JAMA Psychiatry. 2024;81:1142–7.10.1001/jamapsychiatry.2024.2553 39230917

5.Popov T Tröndle M Baranczuk-Turska Z Pfeiffer C Haufe S Langer N Test–retest reliability of resting-state EEG in young and older adults Psychophysiology 2023 60 e14268 36894751 10.1111/psyp.14268 Popov T, Tröndle M, Baranczuk-Turska Z, Pfeiffer C, Haufe S, Langer N. Test–retest reliability of resting-state EEG in young and older adults. Psychophysiology. 2023;60:e14268.10.1111/psyp.14268 36894751

6.Rahul J Sharma D Sharma LD Nanda U Sarkar AK A systematic review of EEG based automated schizophrenia classification through machine learning and deep learning Front Hum Neurosci 2024 18 1347082 38419961 10.3389/fnhum.2024.1347082 PMC10899326 Rahul J, Sharma D, Sharma LD, Nanda U, Sarkar AK. A systematic review of EEG based automated schizophrenia classification through machine learning and deep learning. Front Hum Neurosci. 2024;18:1347082.10.3389/fnhum.2024.1347082 PMC10899326 38419961

7.Saha A Park S Geem ZW Singh PK Schizophrenia detection and classification: a systematic review of the last decade Diagnostics 2024 14 2698 39682605 10.3390/diagnostics14232698 PMC11640697 Saha A, Park S, Geem ZW, Singh PK. Schizophrenia detection and classification: a systematic review of the last decade. Diagnostics. 2024;14:2698.10.3390/diagnostics14232698 PMC11640697 39682605

8.Sun J Cao R Zhou M Hussain W Wang B Xue J A hybrid deep neural network for classification of schizophrenia using EEG Data Sci Rep 2021 11 4706 10.1038/s41598-021-83350-6 33633134 PMC7907145 Sun J, Cao R, Zhou M, Hussain W, Wang B, Xue J, et al. A hybrid deep neural network for classification of schizophrenia using EEG Data. Sci Rep. 2021;11:4706.10.1038/s41598-021-83350-6 PMC7907145 33633134

9.Shoeibi A Sadeghi D Moridian P Ghassemi N Heras J Alizadehsani R Automatic diagnosis of schizophrenia in EEG signals using CNN-LSTM models Front Neuroinform 2021 15 777977 34899226 10.3389/fninf.2021.777977 PMC8657145 Shoeibi A, Sadeghi D, Moridian P, Ghassemi N, Heras J, Alizadehsani R, et al. Automatic diagnosis of schizophrenia in EEG signals using CNN-LSTM models. Front Neuroinform. 2021;15:777977.10.3389/fninf.2021.777977 PMC8657145 34899226

10.Aslan Z Akin M A deep learning approach in automated detection of schizophrenia using scalogram images of EEG signals Phys Eng Sci Med 2022 45 83 96 10.1007/s13246-021-01083-2 34822131 Aslan Z, Akin M. A deep learning approach in automated detection of schizophrenia using scalogram images of EEG signals. Phys Eng Sci Med. 2022;45:83–96.10.1007/s13246-021-01083-2 34822131

11.Bagherzadeh S Shahabi MS Shalbaf A Detection of schizophrenia using hybrid of deep learning and brain effective connectivity image from electroencephalogram signal Comput Biol Med 2022 146 105570 35504218 10.1016/j.compbiomed.2022.105570 Bagherzadeh S, Shahabi MS, Shalbaf A. Detection of schizophrenia using hybrid of deep learning and brain effective connectivity image from electroencephalogram signal. Comput Biol Med. 2022;146:105570.10.1016/j.compbiomed.2022.105570 35504218

12.Ilakiyaselvan N Khan AN Shahina A Reconstructed phase space portraits for detecting brain diseases using deep learning Biomed Signal Process Control 2022 71 103278 10.1016/j.bspc.2021.103278 Ilakiyaselvan N, Khan AN, Shahina A. Reconstructed phase space portraits for detecting brain diseases using deep learning. Biomed Signal Process Control. 2022;71:103278.

13.Sharma G Joshi AM SzHNN: A novel and scalable deep convolution hybrid neural network framework for schizophrenia detection using multichannel EEG IEEE Trans Instrum Meas 2022 71 1 9 Sharma G, Joshi AM. SzHNN: A novel and scalable deep convolution hybrid neural network framework for schizophrenia detection using multichannel EEG. IEEE Trans Instrum Meas. 2022;71:1–9.

14.Urkin B Parnas J Raballo A Koren D Schizophrenia spectrum disorders: an empirical benchmark study of real-world diagnostic accuracy and reliability among leading international psychiatrists Schizophr Bull Open 2024 5 sgae012 39144107 10.1093/schizbullopen/sgae012 PMC11207759 Urkin B, Parnas J, Raballo A, Koren D. Schizophrenia spectrum disorders: an empirical benchmark study of real-world diagnostic accuracy and reliability among leading international psychiatrists. Schizophr Bull Open. 2024;5:sgae012.10.1093/schizbullopen/sgae012 PMC11207759 39144107

15.Sharma G Joshi AM Yadav D Mohanty SP A smart healthcare framework for accurate detection of schizophrenia using multichannel EEG IEEE Trans Instrum Meas 2023 72 1 9 Sharma G, Joshi AM, Yadav D, Mohanty SP. A smart healthcare framework for accurate detection of schizophrenia using multichannel EEG. IEEE Trans Instrum Meas. 2023;72:1–9.

16.Grover N Chharia A Upadhyay R Longo L Schizo-net: a novel schizophrenia diagnosis framework using late fusion multimodal deep learning on electroencephalogram-based brain connectivity indices IEEE Trans Neural Syst Rehabil Eng 2023 31 464 73 37022027 10.1109/TNSRE.2023.3237375 Grover N, Chharia A, Upadhyay R, Longo L. Schizo-net: a novel schizophrenia diagnosis framework using late fusion multimodal deep learning on electroencephalogram-based brain connectivity indices. IEEE Trans Neural Syst Rehabil Eng. 2023;31:464–73.10.1109/TNSRE.2023.3237375 37022027

17.Kapoor S Narayanan A Leakage and the reproducibility crisis in machine-learning-based science Patterns 2023 4 100804 37720327 10.1016/j.patter.2023.100804 PMC10499856 Kapoor S, Narayanan A. Leakage and the reproducibility crisis in machine-learning-based science. Patterns. 2023;4:100804.10.1016/j.patter.2023.100804 PMC10499856 37720327

18.Sasse L Nicolaisen-Sobesky E Dukart J Eickhoff S Götz M Hamdan S Overview of leakage scenarios in supervised machine learning J Big Data 2025 12 135 10.1186/s40537-025-01193-8 Sasse L, Nicolaisen-Sobesky E, Dukart J, Eickhoff S, Götz M, Hamdan S, et al. Overview of leakage scenarios in supervised machine learning. J Big Data. 2025;12:135.

19.Del Pup F Zanola A Tshimanga LF Bertoldo A Finos L Atzori M The role of data partitioning on the performance of EEG-based deep learning models in supervised cross-subject analysis: a preliminary study Comput Biol Med 2025 196 110608 40602315 10.1016/j.compbiomed.2025.110608 Del Pup F, Zanola A, Tshimanga LF, Bertoldo A, Finos L, Atzori M. The role of data partitioning on the performance of EEG-based deep learning models in supervised cross-subject analysis: a preliminary study. Comput Biol Med. 2025;196:110608.10.1016/j.compbiomed.2025.110608 40602315

20.Shim M Lee S-H Hwang H-J Inflated prediction accuracy of neuropsychiatric biomarkers caused by data leakage in feature selection Sci Rep 2021 11 7980 10.1038/s41598-021-87157-3 33846489 PMC8042090 Shim M, Lee S-H, Hwang H-J. Inflated prediction accuracy of neuropsychiatric biomarkers caused by data leakage in feature selection. Sci Rep. 2021;11:7980.10.1038/s41598-021-87157-3 PMC8042090 33846489

21.Kaufman S Rosset S Perlich C Stitelman O Leakage in data mining: formulation, detection, and avoidance ACM Trans Knowl Discov Data 2012 6 1 21 10.1145/2382577.2382579 Kaufman S, Rosset S, Perlich C, Stitelman O. Leakage in data mining: formulation, detection, and avoidance. ACM Trans Knowl Discov Data. 2012;6:1–21.

22.Lee H-T Cheon H-R Lee S-H Shim M Hwang H-J Risk of data leakage in estimating the diagnostic performance of a deep-learning-based computer-aided system for psychiatric disorders Sci Rep 2023 13 16633 10.1038/s41598-023-43542-8 37789047 PMC10547830 Lee H-T, Cheon H-R, Lee S-H, Shim M, Hwang H-J. Risk of data leakage in estimating the diagnostic performance of a deep-learning-based computer-aided system for psychiatric disorders. Sci Rep. 2023;13:16633.10.1038/s41598-023-43542-8 PMC10547830 37789047

23.Starcke J Spadafora J Spadafora J Spadafora P Toma M The effect of data leakage and feature selection on machine learning performance for early parkinson’s disease detection Bioengineering 2025 12 845 40868358 10.3390/bioengineering12080845 PMC12383348 Starcke J, Spadafora J, Spadafora J, Spadafora P, Toma M. The effect of data leakage and feature selection on machine learning performance for early parkinson’s disease detection. Bioengineering. 2025;12:845.10.3390/bioengineering12080845 PMC12383348 40868358

24.Young VM Gates S Garcia LY Salardini A Data leakage in deep learning for alzheimer’s disease diagnosis: a scoping review of methodological rigor and performance inflation Diagnostics 2025 15 2348 41008719 10.3390/diagnostics15182348 PMC12468286 Young VM, Gates S, Garcia LY, Salardini A. Data leakage in deep learning for alzheimer’s disease diagnosis: a scoping review of methodological rigor and performance inflation. Diagnostics. 2025;15:2348.10.3390/diagnostics15182348 PMC12468286 41008719

25.Rosenblatt M Tejavibulya L Jiang R Noble S Scheinost D Data leakage inflates prediction performance in connectome-based machine learning models Nat Commun 2024 15 1829 10.1038/s41467-024-46150-w 38418819 PMC10901797 Rosenblatt M, Tejavibulya L, Jiang R, Noble S, Scheinost D. Data leakage inflates prediction performance in connectome-based machine learning models. Nat Commun. 2024;15:1829.10.1038/s41467-024-46150-w PMC10901797 38418819

26.Brookshire G Kasper J Blauch NM Wu YC Glatt R Merrill DA Data leakage in deep learning studies of translational EEG Front Neurosci 2024 18 1373515 38765672 10.3389/fnins.2024.1373515 PMC11099244 Brookshire G, Kasper J, Blauch NM, Wu YC, Glatt R, Merrill DA, et al. Data leakage in deep learning studies of translational EEG. Front Neurosci. 2024;18:1373515.10.3389/fnins.2024.1373515 PMC11099244 38765672

27.Ranjan R Sahana BC Bhandari AK Deep learning models for diagnosis of schizophrenia using EEG signals: emerging trends, challenges, and prospects Arch Comput Methods Eng 2024 31 2345 84 10.1007/s11831-023-10047-6 Ranjan R, Sahana BC, Bhandari AK. Deep learning models for diagnosis of schizophrenia using EEG signals: emerging trends, challenges, and prospects. Arch Comput Methods Eng. 2024;31:2345–84.

28.Uyanik H Sengur A Salvi M Tan RS Tan JH Acharya UR Automated detection of neurological and mental health disorders using EEG signals and artificial intelligence: a systematic review Wiley Interdiscip Rev: Data Min Knowl Discov 2025 15 e70002 Uyanik H, Sengur A, Salvi M, Tan RS, Tan JH, Acharya UR. Automated detection of neurological and mental health disorders using EEG signals and artificial intelligence: a systematic review. Wiley Interdiscip Rev: Data Min Knowl Discov. 2025;15:e70002.

29.Jiang S Jia Q Peng Z Zhou Q An Z Chen J Can artificial intelligence be the future solution to the enormous challenges and suffering caused by Schizophrenia? Schizophrenia 2025 11 32 40021674 10.1038/s41537-025-00583-4 PMC11871033 Jiang S, Jia Q, Peng Z, Zhou Q, An Z, Chen J, et al. Can artificial intelligence be the future solution to the enormous challenges and suffering caused by Schizophrenia?. Schizophrenia. 2025;11:32.10.1038/s41537-025-00583-4 PMC11871033 40021674

30.Khare SK Bajaj V Acharya UR SchizoNET: a robust and accurate Margenau–Hill time-frequency distribution based deep neural network model for schizophrenia detection using EEG signals Physiol Meas 2023 44 035005 10.1088/1361-6579/acbc06 36787641 Khare SK, Bajaj V, Acharya UR. SchizoNET: a robust and accurate Margenau–Hill time-frequency distribution based deep neural network model for schizophrenia detection using EEG signals. Physiol Meas. 2023;44:035005.10.1088/1361-6579/acbc06 36787641

31.Oh SL Vicnesh J Ciaccio EJ Yuvaraj R Acharya UR Deep convolutional neural network model for automated diagnosis of schizophrenia using EEG signals Appl Sci-Basel 2019 9 2870 10.3390/app9142870 Oh SL, Vicnesh J, Ciaccio EJ, Yuvaraj R, Acharya UR. Deep convolutional neural network model for automated diagnosis of schizophrenia using EEG signals. Appl Sci-Basel. 2019;9:2870.

32.Aydemir E Dogan S Baygin M Ooi CP Barua PD Tuncer T CGP17Pat: automated schizophrenia detection based on a cyclic group of prime order patterns using EEG signals Healthc 2022 10 643 10.3390/healthcare10040643 PMC9027158 35455821 Aydemir E, Dogan S, Baygin M, Ooi CP, Barua PD, Tuncer T, et al. CGP17Pat: automated schizophrenia detection based on a cyclic group of prime order patterns using EEG signals. Healthc. 2022;10:643.10.3390/healthcare10040643 PMC9027158 35455821

33.Hassan F Hussain SF Qaisar SM Fusion of multivariate EEG signals for schizophrenia detection using CNN and machine learning techniques Inf Fusion 2023 92 466 78 10.1016/j.inffus.2022.12.019 Hassan F, Hussain SF, Qaisar SM. Fusion of multivariate EEG signals for schizophrenia detection using CNN and machine learning techniques. Inf Fusion. 2023;92:466–78.

34.Li B Wang J Guo Z Li Y Automatic detection of schizophrenia based on spatial–temporal feature mapping and LeViT with EEG signals Expert Syst Appl 2023 224 119969 10.1016/j.eswa.2023.119969 Li B, Wang J, Guo Z, Li Y. Automatic detection of schizophrenia based on spatial–temporal feature mapping and LeViT with EEG signals. Expert Syst Appl. 2023;224:119969.

35.Borisov S Kaplan AY Gorbachevskaya N Kozlova I Analysis of EEG structural synchrony in adolescents with schizophrenic disorders Hum Physiol 2005 31 255 61 10.1007/s10747-005-0042-z 16004396 Borisov S, Kaplan AY, Gorbachevskaya N, Kozlova I. Analysis of EEG structural synchrony in adolescents with schizophrenic disorders. Hum Physiol. 2005;31:255–61.16004396

36.Olejarczyk E Jernajczyk W Graph-based analysis of brain connectivity in schizophrenia PLoS One 2017 12 e0188629 29190759 10.1371/journal.pone.0188629 PMC5708839 Olejarczyk E, Jernajczyk W. Graph-based analysis of brain connectivity in schizophrenia. PLoS One. 2017;12:e0188629.10.1371/journal.pone.0188629 PMC5708839 29190759

37.Ford JM Palzes VA Roach BJ Mathalon DH Did I do that? Abnormal predictive processes in schizophrenia when button pressing to deliver a tone Schizophr Bull 2014 40 804 12 23754836 10.1093/schbul/sbt072 PMC4059422 Ford JM, Palzes VA, Roach BJ, Mathalon DH. Did I do that? Abnormal predictive processes in schizophrenia when button pressing to deliver a tone. Schizophr Bull. 2014;40:804–12.10.1093/schbul/sbt072 PMC4059422 23754836

38.Racz FS Farkas K Becske M Molnar H Fodor Z Mukli P Reduced temporal variability of cortical excitation/inhibition ratio in schizophrenia Schizophrenia 2025 11 20 39966406 10.1038/s41537-025-00568-3 PMC11836122 Racz FS, Farkas K, Becske M, Molnar H, Fodor Z, Mukli P, et al. Reduced temporal variability of cortical excitation/inhibition ratio in schizophrenia. Schizophrenia. 2025;11:20.10.1038/s41537-025-00568-3 PMC11836122 39966406

39.He K, Zhang X, Ren S, and Sun J. Deep residual learning for image recognition. In Proc. IEEE conference on Computer Vision and Pattern Recognition. 2016;pp. 770–8.

40.Shin H-C Roth HR Gao M Lu L Xu Z Nogues I Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning IEEE Trans Med Imaging 2016 35 1285 98 26886976 10.1109/TMI.2016.2528162 PMC4890616 Shin H-C, Roth HR, Gao M, Lu L, Xu Z, Nogues I, et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning. IEEE Trans Med Imaging. 2016;35:1285–98.10.1109/TMI.2016.2528162 PMC4890616 26886976

41.Chen T, Guestrin C. Xgboost: a scalable tree boosting system. In Proc. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016;pp. 785–94.

42.Urbanowicz RJ Meeker M La Cava W Olson RS Moore JH Relief-based feature selection: introduction and review J Biomed Inform 2018 85 189 203 30031057 10.1016/j.jbi.2018.07.014 PMC6299836 Urbanowicz RJ, Meeker M, La Cava W, Olson RS, Moore JH. Relief-based feature selection: introduction and review. J Biomed Inform. 2018;85:189–203.10.1016/j.jbi.2018.07.014 PMC6299836 30031057

43.Sugden RJ Diamandis P Generalizable electroencephalographic classification of Parkinson’s disease using deep learning Inform Med Unlocked 2023 42 101352 10.1016/j.imu.2023.101352 Sugden RJ, Diamandis P. Generalizable electroencephalographic classification of Parkinson’s disease using deep learning. Inform Med Unlocked. 2023;42:101352.

44.Peng HC Long FH Ding C Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy IEEE Trans Pattern Anal Mach Intell 2005 27 1226 38 16119262 10.1109/TPAMI.2005.159 Peng HC, Long FH, Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell. 2005;27:1226–38.10.1109/TPAMI.2005.159 16119262

45.Kuruppu G Wagh N Kremen V Varatharajah Y EEG foundation models: a critical review of current progress and future directions J Neural Eng 2026 23 021001 10.1088/1741-2552/ae4455 41666566 Kuruppu G, Wagh N, Kremen V, Varatharajah Y. EEG foundation models: a critical review of current progress and future directions. J Neural Eng. 2026;23:021001.10.1088/1741-2552/ae4455 41666566

46.El Ouahidi Y, Lys J, Thölke P, Farrugia N, Pasdeloup B, Gripon V, et al. REVE: A foundation model for EEG-adapting to any setup with large-scale pretraining on 25,000 subjects. Advances in Neural Information Processing Systems. 2026;38:22541–77.

47.van der Maaten L Hinton G Visualizing data using t-SNE J Mach Learn Res 2008 9 2579 605 van der Maaten L, Hinton G. Visualizing data using t-SNE. J Mach Learn Res. 2008;9:2579–605.

48.Raiaan MAK Sakib S Fahad NM Al Mamun A Rahman MA Shatabda S A systematic review of hyperparameter optimization techniques in convolutional neural networks Decis Anal J 2024 11 100470 10.1016/j.dajour.2024.100470 Raiaan MAK, Sakib S, Fahad NM, Al Mamun A, Rahman MA, Shatabda S, et al. A systematic review of hyperparameter optimization techniques in convolutional neural networks. Decis Anal J. 2024;11:100470.

49.Wojciuk M Swiderska-Chadaj Z Siwek K Gertych A Improving classification accuracy of fine-tuned CNN models: impact of hyperparameter optimization Heliyon 2024 10 e26586 38463880 10.1016/j.heliyon.2024.e26586 PMC10920154 Wojciuk M, Swiderska-Chadaj Z, Siwek K, Gertych A. Improving classification accuracy of fine-tuned CNN models: impact of hyperparameter optimization. Heliyon. 2024;10:e26586.10.1016/j.heliyon.2024.e26586 PMC10920154 38463880

50.Khare SK Bajaj V Acharya UR SPWVD-CNN for automated detection of schizophrenia patients using EEG signals IEEE Trans Instrum Meas 2021 70 1 9 10.1109/TIM.2021.3070608 Khare SK, Bajaj V, Acharya UR. SPWVD-CNN for automated detection of schizophrenia patients using EEG signals. IEEE Trans Instrum Meas. 2021;70:1–9.

51.Hosseini M Powell M Collins J Callahan-Flintoft C Jones W Bowman H I tried a bunch of things: The dangers of unexpected overfitting in classification of brain data Neurosci Biobehav Rev 2020 119 456 67 33035522 10.1016/j.neubiorev.2020.09.036 Hosseini M, Powell M, Collins J, Callahan-Flintoft C, Jones W, Bowman H, et al. I tried a bunch of things: The dangers of unexpected overfitting in classification of brain data. Neurosci Biobehav Rev. 2020;119:456–67.10.1016/j.neubiorev.2020.09.036 33035522

52.Supakar R Satvaya P Chakrabarti P A deep learning based model using RNN-LSTM for the detection of schizophrenia from EEG data Comput Biol Med 2022 151 106225 36306576 10.1016/j.compbiomed.2022.106225 Supakar R, Satvaya P, Chakrabarti P. A deep learning based model using RNN-LSTM for the detection of schizophrenia from EEG data. Comput Biol Med. 2022;151:106225.10.1016/j.compbiomed.2022.106225 36306576

53.Shoeibi A, Jafari M, Sadeghi D, Alizadehsani R, Alinejad-Rokny H, Beheshti A, et al. Early diagnosis of schizophrenia in EEG signals using one dimensional transformer model. In International Work-Conference on the Interplay between Natural and Artificial Computation. Cham: Springer Nature Switzerland. 2024;pp. 139–49.

54.Ouzzani M Hammady H Fedorowicz Z Elmagarmid A Rayyan—a web and mobile app for systematic reviews Syst Rev 2016 5 210 10.1186/s13643-016-0384-4 27919275 PMC5139140 Ouzzani M, Hammady H, Fedorowicz Z, Elmagarmid A. Rayyan—a web and mobile app for systematic reviews. Syst Rev. 2016;5:210.10.1186/s13643-016-0384-4 PMC5139140 27919275

55.Gorbachevskaya NN, Borisov SV. EEG of healthy adolescents and adolescents with symptoms of schizophrenia. MV Lomonosov Moscow State University. [Online]. Available: http://brain.bio.msu.ru/eegschizophrenia.htm. 2019.

56.Kulaichev AP Gorbachevskaya NL Differentiation of norm and disorders of schizophrenic spectrum by analysis of EEG correlation synchrony J Exp Integr Med 2013 3 267 10.5455/jeim.060713.or.081 Kulaichev AP, Gorbachevskaya NL. Differentiation of norm and disorders of schizophrenic spectrum by analysis of EEG correlation synchrony. J Exp Integr Med. 2013;3:267.

57.Jasper HH The ten-twenty electrode system of the international federation Electroencephalogr Clin Neurophysiol 1958 10 370 10590970 Jasper HH. The ten-twenty electrode system of the international federation. Electroencephalogr Clin Neurophysiol. 1958;10:370.10590970

58.Delorme A Makeig S EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis J Neurosci Methods 2004 134 9 21 15102499 10.1016/j.jneumeth.2003.10.009 Delorme A, Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Methods. 2004;134:9–21.10.1016/j.jneumeth.2003.10.009 15102499

59.Bigdely-Shamlo N Mullen T Kothe C Su K-M Robbins KA The PREP pipeline: standardized preprocessing for large-scale EEG analysis Front Neuroinform 2015 9 16 26150785 10.3389/fninf.2015.00016 PMC4471356 Bigdely-Shamlo N, Mullen T, Kothe C, Su K-M, Robbins KA. The PREP pipeline: standardized preprocessing for large-scale EEG analysis. Front Neuroinform. 2015;9:16.10.3389/fninf.2015.00016 PMC4471356 26150785

60.Hyvarinen A Oja E Independent component analysis: algorithms and applications Neural Netw 2000 13 411 30 10946390 10.1016/S0893-6080(00)00026-5 Hyvarinen A, Oja E. Independent component analysis: algorithms and applications. Neural Netw. 2000;13:411–30.10.1016/s0893-6080(00)00026-5 10946390

61.Winkler I Brandl S Horn F Waldburger E Allefeld C Tangermann M Robust artifactual independent component classification for BCI practitioners J Neural Eng 2014 11 035013 24836294 10.1088/1741-2560/11/3/035013 Winkler I, Brandl S, Horn F, Waldburger E, Allefeld C, Tangermann M. Robust artifactual independent component classification for BCI practitioners. J Neural Eng. 2014;11:035013.10.1088/1741-2560/11/3/035013 24836294

62.Bhatt D Patel C Talsania H Patel J Vaghela R Pandya S CNN variants for computer vision: history, architecture, application, challenges and future scope Electronics 2021 10 2470 10.3390/electronics10202470 Bhatt D, Patel C, Talsania H, Patel J, Vaghela R, Pandya S, et al. CNN variants for computer vision: history, architecture, application, challenges and future scope. Electronics. 2021;10:2470.

63.Rajwal S Aggarwal S Convolutional neural network-based EEG signal analysis: a systematic review Arch Comput Methods Eng 2023 30 3585 615 10.1007/s11831-023-09920-1 Rajwal S, Aggarwal S. Convolutional neural network-based EEG signal analysis: a systematic review. Arch Comput Methods Eng. 2023;30:3585–615.

64.Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L. Imagenet: a large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition. IEEE. 2009;pp. 248–55.

65.Hosseini M-P Hosseini A Ahi K A review on machine learning for EEG signal processing in bioengineering IEEE Rev Biomed Eng 2020 14 204 18 10.1109/RBME.2020.2969915 32011262 Hosseini M-P, Hosseini A, Ahi K. A review on machine learning for EEG signal processing in bioengineering. IEEE Rev Biomed Eng. 2020;14:204–18.10.1109/RBME.2020.2969915 32011262

66.Santos-Mayo L San-Jose-Revuelta LM Arribas JI A computer-aided diagnosis system with EEG based on the P3b wave during an auditory odd-ball task in schizophrenia IEEE Trans Biomed Eng 2016 64 395 407 10.1109/TBME.2016.2558824 28113193 Santos-Mayo L, San-Jose-Revuelta LM, Arribas JI. A computer-aided diagnosis system with EEG based on the P3b wave during an auditory odd-ball task in schizophrenia. IEEE Trans Biomed Eng. 2016;64:395–407.10.1109/TBME.2016.2558824 28113193

67.Racz FS Stylianou O Mukli P Eke A Multifractal and entropy-based analysis of delta band neural activity reveals altered functional connectivity dynamics in schizophrenia Front Syst Neurosci 2020 14 49 32792917 10.3389/fnsys.2020.00049 PMC7394222 Racz FS, Stylianou O, Mukli P, Eke A. Multifractal and entropy-based analysis of delta band neural activity reveals altered functional connectivity dynamics in schizophrenia. Front Syst Neurosci. 2020;14:49.10.3389/fnsys.2020.00049 PMC7394222 32792917

68.Guo Z, Wu L, Li Y, and Li B. Deep neural network classification of EEG data in schizophrenia. In 2021 IEEE 10th Data Driven Control and Learning Systems Conference (DDCLS). IEEE. 2021;pp. 1322–27.

69.Dimitriadis SI Reconfiguration of αmplitude driven dominant coupling modes (DoCM) mediated by α-band in adolescents with schizophrenia spectrum disorders Prog Neuropsychopharmacol Biol Psychiatry 2021 108 110073 32805332 10.1016/j.pnpbp.2020.110073 Dimitriadis SI. Reconfiguration of αmplitude driven dominant coupling modes (DoCM) mediated by α-band in adolescents with schizophrenia spectrum disorders. Prog Neuropsychopharmacol Biol Psychiatry. 2021;108:110073.10.1016/j.pnpbp.2020.110073 32805332

70.De Miras JR Ibáñez-Molina AJ Soriano MF Iglesias-Parro S Schizophrenia classification using machine learning on resting state EEG signal Biomed Signal Process Control 2023 79 104233 10.1016/j.bspc.2022.104233 De Miras JR, Ibáñez-Molina AJ, Soriano MF, Iglesias-Parro S. Schizophrenia classification using machine learning on resting state EEG signal. Biomed Signal Process Control. 2023;79:104233.

71.Hjorth B EEG analysis based on time domain properties Electroencephalogr Clin Neurophysiol 1970 29 306 10 4195653 10.1016/0013-4694(70)90143-4 Hjorth B. EEG analysis based on time domain properties. Electroencephalogr Clin Neurophysiol. 1970;29:306–10.10.1016/0013-4694(70)90143-4 4195653

72.Márton L Brassai ST Bakó L Losonczi L Detrended fluctuation analysis of EEG signals Procedia Technol 2014 12 125 32 10.1016/j.protcy.2013.12.465 Márton L, Brassai ST, Bakó L, Losonczi L. Detrended fluctuation analysis of EEG signals. Procedia Technol. 2014;12:125–32.

73.Kaposzta Z Czoch A Stylianou O Kim K Mukli P Eke A Real-time algorithm for detrended cross-correlation analysis of long-range coupled processes Front Physiol 2022 13 817268 35360238 10.3389/fphys.2022.817268 PMC8963246 Kaposzta Z, Czoch A, Stylianou O, Kim K, Mukli P, Eke A, et al. Real-time algorithm for detrended cross-correlation analysis of long-range coupled processes. Front Physiol. 2022;13:817268.10.3389/fphys.2022.817268 PMC8963246 35360238

74.Bandt C Pompe B Permutation entropy: a natural complexity measure for time series Phys Rev Lett 2002 88 174102 12005759 10.1103/PhysRevLett.88.174102 Bandt C, Pompe B. Permutation entropy: a natural complexity measure for time series. Phys Rev Lett. 2002;88:174102.10.1103/PhysRevLett.88.174102 12005759

75.Montez T Linkenkaer-Hansen K van Dijk BW Stam CJ Synchronization likelihood with explicit time-frequency priors Neuroimage 2006 33 1117 25 17023181 10.1016/j.neuroimage.2006.06.066 Montez T, Linkenkaer-Hansen K, van Dijk BW, Stam CJ. Synchronization likelihood with explicit time-frequency priors. Neuroimage. 2006;33:1117–25.10.1016/j.neuroimage.2006.06.066 17023181

76.Stam CJ Nolte G Daffertshofer A Phase lag index: assessment of functional connectivity from multi channel EEG and MEG with diminished bias from common sources Hum Brain Mapp 2007 28 1178 93 17266107 10.1002/hbm.20346 PMC6871367 Stam CJ, Nolte G, Daffertshofer A. Phase lag index: assessment of functional connectivity from multi channel EEG and MEG with diminished bias from common sources. Hum Brain Mapp. 2007;28:1178–93.10.1002/hbm.20346 PMC6871367 17266107

77.Rubinov M Sporns O Complex network measures of brain connectivity: uses and interpretations NeuroImage 2010 52 1059 69 19819337 10.1016/j.neuroimage.2009.10.003 Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. NeuroImage. 2010;52:1059–69.10.1016/j.neuroimage.2009.10.003 19819337

78.Bretones CS Parra CR Cascón J Borja AL Sotos JM Automatic identification of schizophrenia employing EEG records analyzed with deep learning algorithms Schizophr Res 2023 261 36 46 10.1016/j.schres.2023.09.010 37690170 Bretones CS, Parra CR, Cascón J, Borja AL, Sotos JM. Automatic identification of schizophrenia employing EEG records analyzed with deep learning algorithms. Schizophr Res. 2023;261:36–46.10.1016/j.schres.2023.09.010 37690170

79.Cohen J A coefficient of agreement for nominal scales Educ Psychol Meas 1960 20 37 46 10.1177/001316446002000104 Cohen J. A coefficient of agreement for nominal scales. Educ Psychol Meas. 1960;20:37–46.

80.Hanley JA McNeil BJ The meaning and use of the area under a receiver operating characteristic (ROC) curve Radiology 1982 143 29 36 7063747 10.1148/radiology.143.1.7063747 Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143:29–36.10.1148/radiology.143.1.7063747 7063747

81.Kumar S Alawieh H Racz FS Fakhreddine R Millán JDR Transfer learning promotes acquisition of individual BCI skills PNAS Nexus 2024 3 pgae076 38426121 10.1093/pnasnexus/pgae076 PMC10903645 Kumar S, Alawieh H, Racz FS, Fakhreddine R, Millán JDR. Transfer learning promotes acquisition of individual BCI skills. PNAS Nexus. 2024;3:pgae076.10.1093/pnasnexus/pgae076 PMC10903645 38426121

82.Moakher M A differential geometric approach to the geometric mean of symmetric positive-definite matrices SIAM J Matrix Anal Appl 2005 26 735 47 10.1137/S0895479803436937 Moakher M. A differential geometric approach to the geometric mean of symmetric positive-definite matrices. SIAM J Matrix Anal Appl. 2005;26:735–47.

83.Ledoit O Wolf M A well-conditioned estimator for large-dimensional covariance matrices J Multivar Anal 2004 88 365 411 10.1016/S0047-259X(03)00096-4 Ledoit O, Wolf M. A well-conditioned estimator for large-dimensional covariance matrices. J Multivar Anal. 2004;88:365–411.

84.Kumar S, Yger F, and Lotte F. Towards adaptive classification using Riemannian geometry approaches in brain-computer interfaces. In 2019 7th International Winter Conference on Brain-Computer Interface (BCI). IEEE. 2019;pp. 1–6.

85.Racz FS Kumar S Kaposzta Z Alawieh H Liu DH Liu R Combining detrended cross-correlation analysis with riemannian geometry-based classification for improved brain-computer interface performance Front Neurosci 2024 18 1271831 38550567 10.3389/fnins.2024.1271831 PMC10973000 Racz FS, Kumar S, Kaposzta Z, Alawieh H, Liu DH, Liu R, et al. Combining detrended cross-correlation analysis with riemannian geometry-based classification for improved brain-computer interface performance. Front Neurosci. 2024;18:1271831.10.3389/fnins.2024.1271831 PMC10973000 38550567

Come citare questo studio

Frigyes Samuel Racz, Gabor Csukly. (2026). Information leakage and performance overestimation in EEG-based schizophrenia detection: evidence from literature and empirical analyses.. Translational psychiatry. DOI: 10.1038/s41398-026-04315-9

Consulta la pubblicazione originale ↗