Idiographic Lapse Prediction With State Space Modeling: Algorithm Development and Validation Study.

Area of Science:

• Mental Health
• Digital Health
• Predictive Modeling

Background:

• Mental health conditions require long-term patient monitoring and personalized treatment.
• Digital sensing and predictive modeling can enhance clinician capacity and personalize care.
• Idiographic approaches, which fit personalized models to individual patients, are of particular interest.

Purpose of the Study:

• To bridge risk prediction and idiographic time-series modeling in mental health.
• To propose state space modeling (SSM) as an alternative to machine learning (ML) classifiers for patient risk prediction.
• To evaluate SSMs against ML classifiers for predicting lapses in alcohol use disorder (AUD).

Main Methods:

• A 3-month observational study of 148 participants in early AUD recovery.
• Idiographic state space models (SSMs) were trained using daily ecological momentary assessment (EMA) data.
• SSM predictive performance was compared to logistic regression and gradient-boosted ML classifiers using AUROC for lapse prediction at 3, 7, and same-day intervals.

Main Results:

• SSMs demonstrated superior mean AUROC performance compared to ML classifiers with 30 or more days of EMA data.
• With sufficient data (≥30 days), SSMs showed high probabilities of best performance for predicting same-day, 3-day, and 7-day lapses.
• Predictive performance varied with limited data (15 days), with SSMs outperforming ML for same-day lapse prediction.

Conclusions:

• SSMs present a compelling alternative to traditional ML for mental health risk prediction.
• SSMs support idiographic model fitting for rare outcomes and offer enhanced predictive performance.
• The SSM framework can be extended beyond risk prediction to optimize treatment selection for various mental health conditions.