Predicting Emotional States Using Behavioral Markers Derived From Passively Sensed Data: Data-Driven Machine Learning...

Area of Science:

• Digital health
• Machine learning
• Mental health

Background:

• Mental health disorders impact mood, cognition, and behavior.
• eHealth and mobile health (mHealth) offer noninvasive data collection for behavioral markers.
• Combining mHealth data with self-reports provides a comprehensive view of mental state.
• mHealth data is often noisy, incomplete, and missing observations, hindering clinical use.

Purpose of the Study:

• Develop a machine learning approach for emotional state prediction.
• Utilize passively collected mobile and wearable device data alongside self-reported emotions.
• Address challenges of high-dimensional, heterogeneous time-series data with missing observations.

Main Methods:

• Analyzed passively sensed behavior and self-reported emotional state data from 943 outpatients.
• Employed probabilistic latent variable models (mixture model and hidden Markov model) for feature extraction.
• Compared classical machine learning methods and recurrent neural networks.
• Proposed a personalized Bayesian model to account for individual differences.

Main Results:

• Probabilistic generative models effectively preprocess and extract features from data with missing observations.
• Models incorporating MM and HMM latent states outperformed others by over 20%.
• Generalized models achieved 0.81 AUC-ROC and 0.71 AUC-PR for predicting emotional valence.
• Personalized models significantly improved prediction performance by considering individual differences.

Conclusions:

• Demonstrated feasibility of machine learning models for predicting emotional states from mobile sensing data.
• Models can effectively handle heterogeneous data with substantial missing observations.
• These models can serve as valuable tools for clinicians monitoring patient mood states.