JMIR Publications

ABSTRACT

Background:

Bipolar disorders (BD) 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 detecThis 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 early deviations indicative of prodromal states.t 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 early deviations indicative of prodromal states.

Methods:

Using high-resolution GPS data from the BipoSense dataset, we applied Density-Based Spatial Clustering of Applications with Noise (DBSCAN) 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 prodromal periods. Optimization of SPC parameters (lambda 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. 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 (e.g., sensitivity >70% and specificity >70%) was achieved for any individual indicator (P > 0.05). However, SPC charts showed temporal patterns potentially useful for future multimodal models.

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. This approach may contribute to early detection of affective episodes in BD and support more timely interventions. Further research is needed to refine these digital biomarkers and validate their clinical utility in reducing the frequency and severity of illness phases.