Differentiating bipolar disorder and schizophrenia using sleep EEG power and coherence features: A machine learning approach based on polysomnography

Differentiating between bipolar disorder (BD) and schizophrenia (SZ) is challenging due to overlapping clinical symptoms and shared genetic risks, resulting in frequent misdiagnoses and ineffective treatments. This study investigates whether overnight polysomnography (PSG), particularly EEG-derived power spectral and coherence features, can robustly distinguish BD from SZ. We collected PSG data from 196 BD and 154 SZ patients and selected a propensity-score matched cohort of 137 patients per group ( N = 274) to obtain comprehensive sleep parameters, EEG power spectra, and coherence metrics. A random forest classifier integrating these features achieved the highest classification accuracy (71.88%), F1-score (0.709), and ROC-AUC (0.770) among tested models, significantly outperforming logistic regression and gradient boosting decision trees. Notably, F3_Theta_Pow, total wake time, C3_Theta_Pow, sleep efficiency emerged as the most discriminative features. Our findings highlight that distinct neurophysiological signatures during sleep effectively differentiate BD from SZ, underscoring the clinical utility of PSG as a practical and objective biomarker. This approach not only addresses diagnostic challenges but also provides insights into the underlying neurobiological mechanisms distinguishing these disorders, potentially guiding more precise clinical interventions.