Depression predictions from GPS-based mobility do not generalize well to large demographically heterogeneous samples

Area of Science:

• Digital Phenotyping
• Machine Learning in Mental Health
• Mobile Sensing Technologies

Background:

• Depression is a prevalent mental health condition impacting millions.
• Mobile sensing and machine learning offer potential for passive depression detection via mobility patterns.
• Previous research often used limited, homogeneous samples like college students.

Purpose of the Study:

• To evaluate the generalizability of mobility-based depression prediction models across diverse populations.
• To assess the impact of socio-demographic heterogeneity on prediction accuracy.
• To determine if training on homogeneous subsamples improves large-scale prediction.

Main Methods:

• Analysis of over 57 million GPS data points from a heterogeneous U.S. sample (N=5,262).
• Comparison of prediction accuracy using three modeling approaches (linear and non-linear).
• Evaluation of model performance on various socio-demographic subgroups.

Main Results:

• A model achieving high accuracy (AUC=0.82) in a homogeneous student sample (N=57) showed only chance-level accuracy (AUC=0.57) in a heterogeneous U.S. sample.
• Low prediction accuracy was observed across different socio-demographic groups.
• Training models on homogeneous subsamples did not significantly improve prediction accuracy.

Conclusions:

• Mobility-based depression prediction models trained on homogeneous samples do not generalize well to diverse populations.
• Significant challenges exist in applying these methods for large-scale, real-world depression detection.
• Further research is needed to develop robust and equitable digital phenotyping tools for mental health.