Interpretable Multimodal Post-Traumatic Stress Disorder Detection via Heterogeneous Graph Attention Networks on Real-World Clinical Data
CMC-COMPUTERS MATERIALS & CONTINUA, cilt.89, sa.2, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 89 Sayı: 2
- Basım Tarihi: 2026
- Doi Numarası: 10.32604/cmc.2026.083509
- Dergi Adı: CMC-COMPUTERS MATERIALS & CONTINUA
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, zbMATH
- İstanbul Üniversitesi-Cerrahpaşa Adresli: Evet
Özet
Objective, interpretable decision support for Post-Traumatic Stress Disorder (PTSD) screening remains a challenge in computational psychiatry, where existing methods either rely on costly neuroimaging or lack the diagnostic transparency required for clinical accountability. This study presents Multimodal HetGAT-PTSD, a heterogeneous graph attention network (HetGAT) that integrates unstructured clinical narratives with structured item-level responses from the PTSD Checklist for DSM-5 (PCL-5). The model operates under a graph topology constrained by the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) criteria to ensure structural alignment between clinical theory and graph-based learning. For each patient, a 25-node directed heterogeneous graph is constructed, connecting a BERTurkencoded narrative node (a pre-trained Transformer-based language model) to 20 PCL-5 symptom nodes and 4 DSM-5 cluster nodes via typed edges that encode clinically grounded relation types. Two stacked graph attention layers with Gated Recurrent Unit (GRU)-style gating propagate multimodal evidence across the graph, and attention-based pooling produces a patient-level representation. The framework was evaluated on 418 real-world Turkish clinical records from a psychiatric training hospital using stratified 5-fold cross-validation. Multimodal HetGAT-PTSD achieved 91.16 +/- 4.25% accuracy, 90.34 +/- 4.58% F1-score, and 94.27 +/- 4.63% AUC-ROC, significantly outperforming the questionnaire-only Multi-Layer Perceptron (MLP) baseline (Delta AUC-ROC = +0.0942, p = 0.020, Cohen's d = 2.532). The model deliberately trades approximately 6% accuracy relative to unconstrained text-only models (p = 0.174, corrected paired t-test) in exchange for mechanistic transparency through its DSM-5-constrained architecture. Explanation stability analysis confirmed perturbation robustness of 0.986 and cross-fold attention similarity of 0.984. The architecture provides a two-tier audit trail comprising edge-level attention weights and node-level pooling scores, enabling clinicians to trace diagnostic evidence from narrative content through PCL-5 items to DSM-5 symptom clusters without post-hoc explanation modules. These preliminary, single-site findings suggest that the framework represents a DSM-5-grounded interpretable research prototype for PTSD decision support that operates on routinely available clinical data, pending external validation on independent, multi-center cohorts.