Abstract Introduction Depression and PTSD are widespread among veterans receiving care for sleep disorders in the U.S. Veterans Affairs system, yet routine screening for Current Major Depressive Episodes (cMDE) remains rare in sleep clinics despite this well-known comorbidity. In this study, we examined how two polysomnography-based machine learning screening pipelines matched clinician-determined cMDE diagnoses in a VA sleep disorders cohort, aiming to address a critical gap in mental health detection. Methods This preliminary analysis included 30 adults recruited at the VA Greater Los Angeles Healthcare System for suspected sleep disorders, with 27 subjects retained after excluding three for data quality issues. cMDE status was established using the Mini International Neuropsychiatric Interview as the reference standard. Two machine learning pipelines were evaluated: one combining six AASM-recommended EEG derivations with a single lead-II ECG, and the other using only the lead-II ECG. Both pipelines performed automated sleep staging and extracted autonomic features, including heart rate and heart rate variability. These physiological features, together with PHQ-9 items 1 and 2, served as inputs to a classifier for estimating cMDE status. Results The prevalence of cMDE was high at 59.3% (16/27). The EEG+ECG pipeline correctly classified 14 of 16 cMDE-positive subjects and 8 of 11 cMDE-negative subjects, resulting in a sensitivity of 87.5% and a specificity of 72.7%. The positive predictive value (PPV) was 82.4% and the negative predictive value (NPV) was 80.0%. The ECG-only pipeline correctly classified 15 of 16 cMDE-positive subjects and 7 of 11 cMDE-negative subjects, with a sensitivity of 93.8% and a specificity of 63.6%. PPV was 78.9% and NPV was 87.5%. Conclusion The exceptionally high prevalence of cMDE in this veteran sample underscores a critical gap in mental health screening within VA sleep clinics, where rates exceed both the general population and many non-VA sleep cohorts. Both polysomnography-based machine learning pipelines evaluated in this preliminary study demonstrated strong screening performance, highlighting the potential for physiology-driven depression screening to be integrated into routine polysomnography workflows and identify veterans who could benefit from timely mental health intervention. Support (if any) Medibio Limited.
Grassi et al. (Fri,) studied this question.