Machine-learning classification of bipolar disorder incorporating wearable-derived core body temperature and actigraphy-derived sleep indices.

Aim: This study aimed to classify patients with bipolar disorder (BD) and normal controls (NCs) using machine-learning models that incorporate wearable-derived core body temperature (CBT) and actigraphy-derived sleep indices. Methods: We enrolled 23 euthymic patients with BD and 43 NCs. Sleep parameters were collected via wrist actigraphy for 14 days, and CBT was measured with a wearable device for 3 days to estimate the CBT nadir and its phase differences with sleep indices. Models were constructed with a base model using CBT- and sleep-derived features, and an extended model that included sociodemographic variables. Features were standardized (StandardScaler), and classifiers (random forest, LightGBM, and XGBoost) were evaluated. Hyperparameters were optimized using cross-validation within the training data. Performance was evaluated using repeated nested cross-validation. SHapley Additive exPlanations (SHAP) values were computed in the extended model to quantify relative feature contributions to the model output. Results: In nested cross-validation at the MaxF1 operating point, the mean area under the receiver operating characteristic curve (ROC-AUC) was 0.771 ± 0.162 for the base model and 0.930 ± 0.050 for the extended model. In the extended model, SHAP suggested that total sleep time, wake time, and the wake time-CBT nadir phase difference were features potentially associated with the model output. Conclusion: In this small-sample study, a classification approach combining wearable-derived CBT indices and actigraphy-based sleep parameters suggested preliminary discrimination between BD and NC. Further validation is warranted in larger cohorts with balanced background characteristics, including disorders requiring differential diagnosis.