EBAS-D: A Deep Learning Framework for High-Precision Stress Detection from ECG Morphology and Heart Rate Variability

Background: Psychological stress contributes significantly to the global burden of mental health disorders, yet current diagnostic methods rely heavily on subjective reporting. Physiological biomarkers, particularly electrocardiogram (ECG) signals, offer promise for objective and scalable stress detection. Methods: We present EBAS-D (ECG-Based AI-enabled Stress Detector), a deep learning model that combines traditional ECG morphology (e.g., RR interval, QTc, QRS duration) with novel ventricular efficiency metrics (PFLVEF and VEI_Model2) and heart rate variability (HRV) analysis. Using data from the PhysioNet spider-fear stress dataset, we introduced a morphology-guided label refinement protocol to distinguish genuine no-stress states from post-stress recovery artifacts. A fully connected neural network was trained on 30-beat ECG segments using weighted binary cross-entropy loss to minimize false positives. Results: EBAS-D was evaluated on 31,466 ECG segments, achieving a specificity of 97.7%, positive predictive value of 99.2%, area under the ROC curve of 0.85, and odds ratio of 43.4 for stress classification. Sensitivity was 50.8% under a conservative decision threshold, prioritizing alarm minimization over recall. The model outperformed baseline classifiers and demonstrated significant differentiation between stress and neutral states (p < 0.0001). Conclusions: This study demonstrates the feasibility of high-specificity, ECG-based stress detection using deep learning and morphology-guided label correction. EBAS-D may augment clinical decision-making in emergency departments, psychiatric triage, or telemedicine, where objective stress assessment can support safe and appropriate care. Further validation across diverse populations and real-world settings is warranted to establish its clinical generalizability.