Deep learning-based stress detection from RR intervals in major depressive disorder, panic disorder, and healthy individuals

Background: Stress exacerbates major depressive disorder (MDD) and panic disorder (PD), highlighting the need for continuous stress quantification. Because stress modulates autonomic function, heart rate variability (HRV) is commonly studied for stress detection. However, conventional HRV pipelines require 5-min recordings and handcrafted features, limiting real-time use. We evaluated whether a one-dimensional (1D) residual network can identify acute cognitive stress directly from ultra-short RR interval (RRI) signals in MDD, PD, and healthy controls (HCs). Methods: One hundred forty-seven adults (MDD = 41, PD = 47, HC = 59) completed up to five lab visits over 12 weeks. At each visit, RRIs were recorded during a 5-min resting baseline and a 5-min mental-arithmetic stressor. A 1D ResNet34 classified baseline versus stress from raw RRIs using both 5-min segments and 1-min epochs. Group-specific models were compared with a combined model trained on pooled data. Generalized estimating equations tested group and phase effects on RRIs. Results: Stress shortened RRIs in every group, but less in patients with MDD and PD than in HC. Combined training outperformed group-specific training: for 5-min data, accuracies reached 0.866 (MDD), 0.865 (PD), and 0.897 (HC); 1-min accuracies were 0.788, 0.815, and 0.797, respectively. Conclusion: Deep learning on raw RRIs detects acute cognitive stress across psychiatric and healthy cohorts without feature engineering. Five-minute windows still yield the best performance, yet 1-min epochs still achieve accuracies of approximately 0.80, demonstrating feasibility for integration into real-time monitoring tools for relapse prevention and personalized care in psychiatry.