Multimodal digital phenotyping of bipolar disorder subtypes: Differences between bipolar disorder I and II based on...

Distinguishing between bipolar disorder type I and II constitutes a significant clinical challenge that relies on retrospective patient recall. Misclassification carries risks of inappropriate pharmacological management. Digital phenotyping offers a potential objective solution by monitoring behavioral biomarkers via smartphones. This study evaluated whether passive sensing data alone, or a multimodal approach combining active and passive data, could distinguish between bipolar disorder type I and II at the population level versus the individual level.

Analyses were conducted on data from 156 patients with newly diagnosed bipolar disorder participating in the Aspirin Bipolar RCT. Passive sensor data from smartphones, including GPS mobility and accelerometer activity metrics, were collected over a six- to twelve-month period. Linear Mixed-Effects Models were employed to analyze population-level differences in behavioral features, controlling for age, sex, and treatment allocation. Machine learning performance was assessed using Explainable Boosting Machines in both generalized (across-patient) and personalized (within-patient) paradigms.

No statistically significant differences in passive behavioral features were found between bipolar disorder type I and II after accounting for multiple testing. However, demographic factors were predictors of behavior: older age was associated with lower sleep fragmentation (p_adj < 0.001), and female sex was associated with a higher proportion of time still (p_adj = 0.007). Consistent with these findings, generalized machine learning models failed to outperform chance at classifying bipolar disorder type (AUC ≈ 0.53). In contrast, personalized models trained on individual data achieved higher discriminative power. A personalized model relying solely on passive sensing achieved an AUC of 0.77, whereas a multimodal model incorporating passive data, active self-reports, and demographic factors reached an AUC of 0.96.

Personalized multimodal digital phenotyping models achieved high performance in distinguishing between bipolar I and II, demonstrating the potential of smartphone-based passive sensing to support diagnostic differentiation at the individual level. However, passive sensing alone yielded substantially lower predictive power. This discriminative power did not generalize across populations, suggesting the absence of a universal digital fingerprint distinguishing bipolar subtypes at the population level. Behavioral markers in bipolar disorder appear to be highly idiosyncratic and strongly influenced by demographic variables. Future clinical applications of digital phenotyping should prioritize personalized modeling approaches that account for individual behavioral baselines.