Deep Learning Analysis of Trauma-Related Emotional Patterns in Couples Facing PTSD

Abstract

This study aimed to examine and classify trauma-related emotional patterns in couples facing PTSD using multimodal deep learning models to predict high-risk emotional episodes during conflict interactions. The study employed a cross-sectional design involving 112 couples (224 individuals) from the United States in which one partner met diagnostic criteria for PTSD. Each couple participated in a 40-minute conflict-discussion task recorded through synchronized audio, video, and physiological sensors. Linguistic transcripts, facial expressions, acoustic features, and autonomic indicators (electrodermal activity, heart-rate variability, and peripheral temperature) were extracted and temporally aligned. These modalities were analyzed using a multimodal transformer architecture integrating text embeddings, CNN-LSTM visual features, physiological time-series data, and acoustic spectral representations. Additional analyses included t-SNE–based latent clustering of emotional patterns and risk-surface modeling of predicted high-risk episodes as a nonlinear function of physiological arousal and negative emotional language. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC metrics with ten-fold stratified cross-validation. The multimodal transformer significantly outperformed baseline and unimodal models in predicting high-risk PTSD-related emotional episodes (AUC = .90; F1 = .83; accuracy = .86; p < .001 vs. clinical baseline). All unimodal models performed above chance but remained significantly weaker than the multimodal approach (p < .01). Latent clustering revealed four statistically distinct emotional interaction patterns (hyperaroused escalation, avoidant disengagement, mixed volatile–repairing, and numbed coexistence), each differing significantly in PTSD severity, relationship quality, and proportion of high-risk segments (all ps < .05). Nonlinear risk-surface analysis demonstrated a strong interaction effect between negative emotional language and physiological arousal in predicting emotional risk (βinteraction = .41, p < .001). Multimodal deep learning provides a highly sensitive and integrative method for identifying trauma-related emotional risk states in couples facing PTSD, revealing clinically meaningful emotional clusters and nonlinear escalation patterns that may inform assessment and intervention.