[[DP - Geolocation]]
[[Mania]]
[[Depression]]
[[Paper - Classifying and clustering mood disorder patients using smartphone data from a feasibility study]]
- # Passive Sensor Data Collected
The mindLAMP app continuously gathered three types of passive smartphone data to capture daily behaviors and routines:
- Geolocation: real-time GPS points to map movement patterns and travel radius
- Accelerometer: device motion readings to estimate physical activity and infer sleep periods
- Screen-state: on/off logs to measure total phone usage and detect routine breaks
---
# Behavioral Features Derived
Raw sensor streams were processed into four key metrics for each 24-hour period:
- Home time: total hours spent at the primary location each day
- Entropy: a quantified measure of how often and widely locations changed, indicating routine stability
- Sleep duration: inferred from sustained periods of phone inactivity, aided by acceleration and screen data
- Screen duration: sum of daily screen-on intervals, reflecting engagement and possible restlessness
---
# How Mood Disorders Were Detected
Researchers transformed these features into input variables for machine learning models:
1. Binary classification
- Distinguished healthy controls from mood-disorder participants using a random forest model
- Passive data alone achieved an AUC of 0.65 for control vs. non-control classification
2. Depression vs. bipolar classification
- Logistic regression on passive data had limited performance (AUC = 0.52)
- Adding active survey responses (PHQ-2, GAD-2) improved AUC to 0.62, highlighting the value of combining passive and active data
3. Unsupervised clustering
- K-means grouped participants into four clusters based on all features
- Moderate silhouette (0.46) and ARI (0.27) scores suggested behavioral patterns partly aligned with clinical diagnoses
These methods show that passively collected smartphone metrics can offer a low-cost, scalable digital biomarker for mood disorders. Further validation with larger, richer datasets is needed before clinical deployment.