Abstract
Post-traumatic stress disorder (PTSD) and post-concussion syndrome comorbid with PTSD (PCS+PTSD) share overlapping symptoms, complicating differential diagnosis reliant on subjective assessments. This research-in-progress evaluates deep learning and interpretable feature-based machine learning models on resting-state functional magnetic resonance imaging (rs-fMRI) data from 87 combat veterans (28 controls, 17 PTSD, 42 PCS+PTSD). Using ROI-based time series (125 regions via cc200 parcellation, 1000 time points per run), InceptionTime achieved 93% accuracy, 92% macro F1-score, and 98% AUC in 5-fold cross-validation, outperforming a baseline CNN (85% accuracy, 84% F1, 94% AUC). Feature-based XGBoost models performed competitively: functional connectivity (FC) yielded 88.6% accuracy and 83.2% macro F1, while combined features reached 88.6% accuracy and 84.7% macro F1. Deep models capture latent temporal dynamics, whereas FC-based approaches provide interpretable insights into network disruptions (e.g., DMN-SN connectivity). Limitations include modest sample size and male-only cohort. Future directions involve external validation on ABIDE and ADHD-200 datasets, temporal attention mechanisms, and dynamic connectivity analysis.
Recommended Citation
Joghataee, Mohammad; Gupta, Ashish; Naeimaeimousavi, Amir; Deshpande, Gopikrishna; and Qin, Xiao, "Multimodal Approaches for Classifying PTSD and PCS-PTSD from Resting-State fMRI: Deep Learning on Time-Series and Interpretable Feature-Based Models" (2025). Proceedings of the 2025 Pre-ICIS SIGDSA Symposium. 8.
https://aisel.aisnet.org/sigdsa2025/8