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

Emotional state prediction and forecasting is a complex task, but it would be a very beneficial tool for detecting early warning signs in clinical treatment. By detecting the risk of relapse of major affective episodes could help catch the early onset of major depressive or manic phases that can be addressed and handled early, leading to reduced severity of symptoms and degree of treatment. 

In this work, we worked on a generic machine learning-based approach for emotional state prediction using passively-collected data from mobile phones and wearable devices, and self-reported emotions by patients. 

We applied probabilistic latent variable models (Mixture Model (MM) and Hidden Markov Model (HMM)) for data averaging and feature extraction on the regularly sampled, but frequently missing and heterogeneous time series data. The extracted features were then combined with a classifier to provide emotional state predictions. Furthermore, we proposed a personalised Bayesian model to improve the performance, which considers the individual differences in the data by applying a different classifier bias term for each patient.