Schizophrenia Detection from EEG Signals: A Transformer Framework with Spectrogram Representation
Abtin Shafiei, Mohsen Hooshmand, Majid Ramezani
cs.AI
Quick verdict
Authors report 92.88% AUC-ROC (CST-SZ) for spectrogram-based schizophrenia detection; leakage-aware protocol sound, but one small split, no significance tests.
Summary
The authors present an EEG-based schizophrenia classification framework built on the public ASZED-153 dataset (153 subjects: 76 patients, 77 controls; 16-channel resting-state recordings at 200/256 Hz). Preprocessed signals (resampling to 256 Hz, 0.5-45 Hz band-pass, 50/100 Hz notch, common average reference, amplitude-based artifact rejection) are converted into channel-averaged STFT spectrograms (Hann window 256, 75% overlap, 128x128 PNGs), then classified with SVM, Random Forest, and XGBoost on 438 handcrafted features, and with three deep models: a four-block CNN (C-SZ), a CNN-Transformer hybrid (CT-SZ), and a Squeeze-and-Excitation variant (CST-SZ). All splits are made at the subject level (70/15/15, stratified) and hyperparameter search uses GroupKFold. On the test set, CST-SZ achieves the best AUC-ROC (92.88%) and AUPRC (0.8209), CT-SZ the best F1 (0.8026), ahead of a reimplemented DWT-CNN baseline (AUC 82.40%) and Random Forest (90.04%). Strengths are the leakage-aware protocol and the relatively large public cohort. Main limitations: a single split with only ~23 test subjects, no confidence intervals, significance tests, or repeated runs; unstated PCA fitting scope; unaddressed medication and acquisition-site confounds; no code release. The margin over Random Forest is modest and statistically unverified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Subject-level partitioning of train/validation/test sets eliminates subject-related data leakage in EEG-based schizophrenia classification.
- STFT spectrogram representations of preprocessed resting-state EEG support competitive schizophrenia detection with both classical ML and deep models.
- The proposed CST-SZ model achieves the best AUC-ROC of 92.88% and AUPRC of 0.8209 on the independent test set, while CT-SZ achieves the best F1-score of 0.8026 (Table III).
- Under the same subject-level protocol, the proposed Transformer family outperforms a reimplemented DWT-CNN baseline (AUC 0.8240) and conventional ML models (best AUC 0.9004).
Strengths
Leakage-aware evaluation design
Sec. IV: stratified 70/15/15 split at subject level, and hyperparameter search for SVM/XGBoost uses 5-fold GroupKFold keyed on subject IDs — directly addressing the sample-level-split pathology the authors correctly diagnose in the EEG-classification literature.
Larger, public cohort than most prior work
ASZED-153 (Sec. III.A, Table I: 153 subjects, 76/77, 16 channels) substantially exceeds the cohorts of the cited baselines, e.g. IBIB PAN (28 subjects) and Moscow (84 subjects) mentioned in the related-work discussion of [8].
Reimplemented external baseline under identical protocol
Complete architecture specification
Weaknesses
Single split, tiny test set, zero uncertainty quantification
Channel averaging discards all spatial information
Reproducibility gaps in the ML pipeline
Colormap indirection unexamined
Confound and site balance not discussed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This preprint proposes a pipeline for automated schizophrenia detection from resting-state EEG on the public ASZED-153 dataset (Sec. III.A, Table I: 153 subjects, 76 patients / 77 controls, 16 channels, two acquisition systems at 200/256 Hz). Recordings pass a preprocessing chain (Sec. III.B: resampling to 256 Hz, 0.5-45 Hz band-pass, 50/100 Hz notch, common average reference, artifact rejection at ±200 μV), are converted to STFT spectrograms averaged over all 16 channels (Sec. III.C: Hann window 256, 75% overlap, dB clipped to [-20, 40], 128x128 PNGs, 4-15 images per subject), and are classified by (a) SVM / Random Forest / XGBoost on a 438-dimensional handcrafted feature vector (Sec. III.E, Table II) and (b) three deep models: C-SZ (four-block CNN), CT-SZ (CNN stem, 256 tokens of dim 128, 4-block Transformer encoder with 4 heads), and CST-SZ (CT-SZ plus Squeeze-and-Excitation blocks). The central claims are that subject-level partitioning removes leakage and that the Transformer family is superior: CST-SZ reaches 92.88% AUC-ROC / 0.8209 AUPRC and CT-SZ the best F1 of 0.8026 (Table III).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What performance does this EEG schizophrenia detection method achieve?
On a 15% held-out test split of the ASZED-153 dataset (153 subjects), the best model, CST-SZ (CNN + Squeeze-and-Excitation + Transformer), reaches AUC-ROC of 92.88% and AUPRC of 0.8209; the CNN-Transformer (CT-SZ) achieves the best F1-score of 0.8026. These are single point estimates from one subject-level split, with no confidence intervals reported.
What dataset does the paper use and is it publicly available?
It uses ASZED-153, an open-access resting-state EEG dataset collected at two institutions in Nigeria: 153 subjects (76 schizophrenia patients, 77 healthy controls), 16 EEG channels, recorded at 200 Hz and 256 Hz with two different acquisition systems.
How does the paper prevent data leakage between training and test data?
All train/validation/test partitions are made at the subject level (70%/15%/15%, stratified), so no recordings from the same person appear in more than one subset, and hyperparameter tuning uses 5-fold GroupKFold grouped by subject identifier. The authors note this is stricter than the sample-level splits common in earlier EEG classification studies.
Is code available for this paper?
The paper does not mention a code release. The dataset is public, and the architectures are described in detail, but training hyperparameters such as learning rate, batch size, and epoch count are not reported, so independent reproduction would require re-engineering parts of the pipeline.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 (framework diagram, demographics, sample spectrograms, CT-SZ architecture, AUPR/AUC overlay) are images and could not be analyzed from the extracted text.
- Only aggregate metrics from Table III are available; no confusion matrices, sensitivity/specificity, or per-subject results appear in the text.
- Some training hyperparameters (learning rate, batch size, epochs) are absent from the extracted text; they may be genuinely unreported rather than lost in extraction.
- This is an arXiv v1 preprint; no peer-review status is stated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.