Emotion Forecasting: A Transformer-Based Approach.

Area of Science:

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

Background:

• Psychiatric patient monitoring is challenging due to subjective assessments and environmental influences.
• Real-time monitoring is crucial for managing the variability of mental states in psychiatric disorders.

Purpose of the Study:

• To develop objective, real-time patient monitoring using deep learning and mobile device data.
• To predict patient self-reports and detect sudden emotional valence changes for timely clinical intervention.

Main Methods:

• Utilized the Evidence-Based Behavior (eB2) app for passive and self-reported data collection.
• Applied hidden Markov models (HMMs) for missing data and transformer deep neural networks for time-series forecasting.
• Employed classification algorithms to predict emotional state and Patient Health Questionnaire-9 (PHQ-9) responses.

Main Results:

• Achieved high accuracy (0.93) and ROC AUC (0.98) for emotional valence classification.
• Successfully predicted emotional state changes one day in advance (ROC AUC 0.87).
• Demonstrated strong predictive performance for PHQ-9 responses, including suicidal ideation (Q9: accuracy 0.9, ROC AUC 0.77).

Conclusions:

• Combined HMM preprocessing with transformer models for stable multivariate time-series forecasting, outperforming traditional methods.
• Showcased the potential of passive variables for predicting patient emotional states and questionnaire scores.
• Enables real-time monitoring for improved risk detection and treatment adjustment in psychiatric care.