Diagnostic Accuracy of Convolutional Neural Networks for Schizophrenia: A Systematic Review and Meta-Analysis.
I. D. Hernández, E. Saucedo, A. Gogeascoechea, D. E. Guajardo, R. Huereca, A. L. Ponce, X. Triana, D. MoralesIntroduction
Schizophrenia remains a major diagnostic challenge due to its clinical heterogeneity and the lack of objective biomarkers. Deep learning (DL), particularly convolutional neural networks (CNNs), enables automated detection of complex neurobiological patterns from neuroimaging data. This study systematically reviewed and meta-analyzed the diagnostic accuracy of CNN-based models applied to electroencephalography (EEG), functional magnetic resonance imaging (fMRI), and structural MRI (sMRI) for schizophrenia classification.
Objectives
To quantify pooled sensitivity and specificity of CNN-based deep learning models for schizophrenia diagnosis across EEG, fMRI, and sMRI modalities, and to evaluate heterogeneity and methodological quality among studies.
Methods
A systematic search of six databases (Medline, Embase, Web of Science, Cochrane Central, PsycInfo, Scopus) was conducted in July 2024, following PRISMA 2020 guidelines (PROSPERO: CRD42024550691). Eligible studies applied CNN architectures to adult schizophrenia cohorts diagnosed per DSM or ICD criteria, reporting diagnostic accuracy metrics. Two reviewers independently extracted data and assessed quality using a modified QUADAS-2 tool tailored for AI/ML studies. Random-effects meta-analyses were performed to pool sensitivity and specificity; summary receiver operating characteristic (SROC) curves were generated for each modality.
Results
Thirty-five studies (n≈13,100 participants) met inclusion criteria. CNN models achieved a pooled sensitivity of 0.90 [95% CI 0.85–0.94] and specificity of 0.90 [0.84–0.93]. EEG-based models reached the highest AUC (0.94) (Figure 1), followed by sMRI (0.91) (Figure 2) and fMRI (0.88) (Figure 3). Multimodal models combining imaging types outperformed unimodal approaches (AUC > 0.95). Moderate heterogeneity (I²≈40%) was observed across studies.