Personalized machine learning of depressed mood using wearables

Area of Science:

• Digital Medicine
• Computational Psychiatry
• Personalized Medicine

Background:

• Depression exhibits significant variability in treatment response.
• Digital medicine and precision therapeutics necessitate personalized approaches for depression management.

Purpose of the Study:

• To develop a systematic pipeline for N-of-1 personalized modeling of depression.
• To generate individualized predictions of depressed mood using multimodal data.
• To identify unique determinants of depression for each individual.

Main Methods:

• Longitudinal ecological momentary assessments, neurocognitive sampling with electroencephalography, and wearable lifestyle data were collected.
• Seven supervised machine learning (ML) approaches were integrated for individual modeling.
• Fourfold nested cross-validation and Shapley statistics were used for model verification and feature importance analysis.

Main Results:

• Individually selected best-fit ML models demonstrated significantly lower prediction errors compared to a general voting regressor.
• No single ML model type was optimal for all individuals, highlighting the need for personalized strategies.
• Shapley values identified distinct, person-specific predictors of depression, including anxiety, exercise, diet, stress, sleep, and neurocognition.

Conclusions:

• A systematic pipeline for N-of-1 personalized depression modeling was established.
• Personalized ML models can accurately predict depressed mood by incorporating individual data.
• Identified personalized features offer potential targets for future ML-guided, multimodal depression treatment strategies.