Using machine learning with intensive longitudinal data to predict depression and suicidal ideation among medical...

Area of Science:

• Digital mental health
• Computational psychiatry
• Machine learning in healthcare

Background:

• Intensive longitudinal methods like ecological momentary assessment and passive sensing are increasingly used with machine learning (ML) to predict depression and suicide risk.
• Existing studies show variability in length, ML methods, and data sources, necessitating research into optimal data collection strategies.
• This study investigated how predictive accuracy for depression and suicidal ideation (SI) changes over time, comparing different ML methods and data sources.

Purpose of the Study:

• To examine the predictive accuracy of ML models for depression and SI over time.
• To compare the performance of different ML algorithms (elastic net regression, random forest) and data sources (daily mood, passive sensing).
• To determine the optimal duration of data collection for accurate mental health risk prediction.

Main Methods:

• Utilized data from 2459 first-year physicians using wearable devices and daily mood assessments.
• Employed linear (elastic net regression) and non-linear (random forest) ML models to predict depression and SI.
• Assessed predictive accuracy iteratively over 92 days, comparing models using daily mood features alone versus mood plus passive-sensing features.

Main Results:

• Elastic net regression models using only daily mood features demonstrated the highest accuracy for predicting mental health outcomes.
• Accurate prediction (within 1 standard error of full 92-day models) was achieved by weeks 7-8.
• Depression at 92 days was accurately predicted (AUC >0.70) after just 14 days of data collection.

Conclusions:

• Simpler ML methods, like elastic net regression with daily mood data, may be more effective than complex methods until passive-sensing features are better defined.
• Intensive longitudinal studies may not require data collection beyond 2 months for significant predictive value in mental health risk assessment.