iSurvive: An Interpretable, Event-time Prediction Model for mHealth

Area of Science:

• Computational Science
• Health Informatics
• Biostatistics

Background:

• Mobile health (mHealth) research increasingly utilizes multimodal data (sensor streams, self-reports) for predictive modeling.
• Interpretable prediction models are vital for domain scientist adoption and informing intervention strategies in mHealth.
• Existing continuous time hidden Markov models (CT-HMMs) can model event data but are not optimized for time-to-event predictions crucial for mHealth.

Purpose of the Study:

• To extend classical survival analysis within a CT-HMM framework for interpretable time-to-event predictions in mHealth.
• To develop a CT-HMM with enhanced emission models capable of handling multimodal data and incorporating domain knowledge.
• To create a model, iSurvive, that addresses the limitations of standard CT-HMMs for predicting future events like substance use relapse.

Main Methods:

• Developed iSurvive, an extension of survival analysis integrated with CT-HMMs.
• Implemented a parameter learning method for generalized linear model (GLM) emissions and survival model fitting.
• Utilized multimodal data, including sensor streams and self-reports, for model training and validation.

Main Results:

• Demonstrated promising predictive performance on both synthetic datasets and a real-world mHealth drug use dataset.
• Showcased the interpretability of the iSurvive model, allowing insights into factors influencing time-to-event predictions.
• Validated the model's ability to handle irregular event data and multimodal inputs effectively.

Conclusions:

• iSurvive offers a powerful and interpretable approach for time-to-event prediction in mHealth using multimodal data.
• The developed methods advance the application of CT-HMMs for complex health prediction tasks.
• This work provides a foundation for improved intervention design and theoretical understanding in mHealth addiction research.