A random forest classifier integrating overnight polysomnography EEG power spectral and coherence features achieved 71.88% accuracy and ROC-AUC of 0.770 in distinguishing bipolar disorder from schizophrenia, with F3_Theta_Pow, total wake time, C3_Theta_Pow, and sleep efficiency emerging as the most discriminative features.
A random forest classifier using PSG-derived EEG features achieved the highest classification performance among tested models for differentiating BD from SZ.
The four most discriminative features for distinguishing bipolar disorder from schizophrenia were F3_Theta_Pow, total wake time, C3_Theta_Pow, and sleep efficiency.
The study used a propensity-score matched cohort to control for confounding variables between the BD and SZ patient groups.
PSG-derived EEG features included both power spectral and coherence metrics collected across an overnight recording.
The study found distinct neurophysiological signatures during sleep that effectively differentiate bipolar disorder from schizophrenia, supporting the clinical utility of PSG as an objective biomarker.
Bipolar disorder and schizophrenia are two serious mental health conditions that are frequently confused with one another because they share many symptoms and even some genetic risk factors. This research suggests that measuring brain activity during sleep using a technique called polysomnography (PSG) — which records brain waves, eye movements, and other signals overnight — can help tell these two conditions apart. By analyzing brain wave patterns and how different brain regions communicate during sleep, researchers were able to build a computer model that correctly classified patients about 72% of the time. The study collected overnight sleep recordings from 196 people with bipolar disorder and 154 people with schizophrenia, then carefully matched 137 patients from each group to make fair comparisons. Among several computer-based classification methods tested, a 'random forest' algorithm performed best, achieving an accuracy of 71.88% and an area under the ROC curve of 0.770. The features that were most useful for distinguishing the two conditions included theta-frequency brain wave power recorded at frontal and central scalp locations, as well as basic sleep quality measures like total time spent awake and overall sleep efficiency. This research suggests that sleep brain activity may carry distinct biological 'fingerprints' for bipolar disorder versus schizophrenia, even when daytime symptoms overlap considerably. If validated in larger studies, this approach could offer clinicians a more objective tool to support diagnosis, potentially reducing the misdiagnoses that currently lead to patients receiving ineffective treatments for extended periods.
Zhong Y, Xu C, Gao Y, Ma H, Liu Y, Yan W, et al.. (2026). Differentiating bipolar disorder and schizophrenia using sleep EEG power and coherence features: A machine learning approach based on polysomnography.. Journal of affective disorders. https://doi.org/10.1016/j.jad.2026.121478