This page was translated automatically from Turkish. Read the original

Bu sayfa, 26. Dünya Psikiyatri Kongresi'nde sunulan çalışmanın yapay zekâ desteğiyle hazırlanmış Türkçe özetidir. Çalışmanın sahibi sunan yazarlardır; atıf için orijinal kaynağa başvurun.

EV1257

Multimodal artificial intelligence for predicting comorbid REM sleep behavior disorder in major depressive disorder

MULTIMODAL AI FOR PREDICTING COMORBID REM SLEEP BEHAVIOR DISORDER IN MAJOR DEPRESSIVE DISORDER

Facial and vocal dynamics extracted from brief speech tasks recorded on a smartphone can distinguish comorbid REM sleep behavior disorder in major depressive disorder with an AUROC of 0.829.

The aim was to test whether facial and vocal dynamics extracted from brief smartphone recordings could distinguish patients with major depressive disorder (MDD) and comorbid REM sleep behavior disorder (RBD) confirmed by video polysomnography (v-PSG) from matched patients with MDD alone, and identify interpretable candidate digital biomarkers. The analysis included 31 participants from a clinical cohort of 329 individuals (19 with MDD and RBD, 12 MDD-only controls), using 261 ten-second clips; the visual pathway processed HOG and facial action unit (AU) features using an sLSTM, while the auditory pathway processed OpenSMILE-based acoustic features using an MLP. Late fusion generated participant-level probabilities, and participant-level five-fold cross-validation was performed to prevent clip-level data leakage. The full model achieved an AUROC of 0.829, an accuracy of 80.5%, a sensitivity (recall) of 90.0%, and an F1 score of 0.848, with a reported mean specificity of 63.3%. In unimodal analyses, audio provided the strongest signal (accuracy 74.9%, F1 0.696); HOG alone yielded 63.4% and 0.594, facial action units alone yielded 70.2% and 0.650, and the combination of action units and audio yielded 77.4% and 0.738, with the full multimodal model achieving the best performance. The model-predicted probability of comorbid RBD was significantly correlated with RBDQ scores (r = 0.53, p = 0.005), showed a trend-level correlation with UPDRS scores (r = 0.34, p = 0.067), and showed no clear association with BDI or HADS scores. In external validation using the Guangzhou cohort (n = 26), which differed in site, language, device, and protocol, audio alone achieved an accuracy of 77.7% and an F1 score of 0.758, while the combination of HOG and action units yielded 69.3% and 0.674. SHAP analysis highlighted lower facial and perioral tension and variability (AU14, AU17, AU25, AU04), along with acoustic measures such as pitch stability, MFCC-based timbre, and voiced/unvoiced segment duration; these were emphasized as candidate biomarkers rather than causal mechanisms. In conclusion, the method is intended not to replace diagnosis but to prioritize targeted RBD assessment in patients with MDD; limitations include the small, cross-sectional sample, the absence of a comparison group with other sleep disorders, and residual medication-related and clinical confounding.

Tüm WCP 2026 özetleri