Predicting Affective Episodes in Bipolar Disorder Using Statistical Process Control Analysis of GPS-Based Mobility Patterns: Quantitative Study.

Background: Bipolar disorders (BDs) represent a significant global health challenge, with frequent and severe affective episodes that impair quality of life. Accurate, early prediction of these episodes remains difficult. Recent advances in mobile sensing offer new possibilities to detect prodromal changes via smart digital phenotypes, such as geolocation data. Objective: This study aimed to examine whether spatial exploratory behavior, assessed via passive GPS data, can predict depressive and manic episodes in individuals with BD. Specifically, we evaluated the predictive value of unique places visited and related mobility metrics using statistical process control (SPC) techniques to identify both early deviations indicative of prodromal states and changes occurring during ongoing affective episodes. Methods: Using high-resolution GPS data from the BipoSense dataset, we applied Density-Based Spatial Clustering of Applications with Noise to extract behavioral mobility indicators: number of unique places visited, frequency of location changes, and time spent per location. We implemented exponentially weighted moving average (EWMA)-based SPC to identify "out-of-bounds" deviations from individual baselines. We then tested the alignment of these deviations with affective episodes and the prodromal periods. Optimization of SPC parameters (λ and control limit L) was performed to enhance predictive accuracy. Results: The analysis included 28 participants with BD and a total of 10,213 observation days, covering 26 depressive and 20 (hypo)manic episodes. Examining whether control limits distinguish affective episodes from euthymic days via multilevel models revealed that median time spent at clusters indicated both depressive and (hypo)manic episodes the best, whereas the number of unique clusters showed no significant association with phase transitions. While EWMA-SPC detected behavioral deviations during affective episodes, no single variable consistently met predefined thresholds for both sensitivity and specificity. Optimized SPC settings improved performance, but the number of unique places alone did not robustly predict prodromal or acute episodes. No statistically significant predictive accuracy (eg, sensitivity >70% and specificity >70%) was achieved for any individual indicator (P>.05). However, some SPC charts suggested within-person temporal deviations preceding episodes, indicating limited yet potentially informative patterns. Conclusions: Although unique places visited alone may not suffice as a predictive marker, the application of EWMA-based SPC to GPS data holds promise for the development of smart digital phenotypes. Although our analysis to predict upcoming episodes did not yield robust predictive accuracy in its current form, it provides a promising conceptual framework for individualized, low-burden behavioral monitoring. Further research is needed to refine existing digital biomarkers, develop new ones, and validate their clinical utility in reducing the frequency and severity of illness phases.