Personalised depression forecasting using mobile sensor data and ecological momentary assessment.

Area of Science:

• Digital Health
• Machine Learning
• Psychiatry

Background:

• Digital health interventions show promise for depression treatment, yet dynamic symptom evolution remains poorly understood.
• Current forecasting models often lack individual patient specificity and overlook fairness across subgroups.
• Personalized models are crucial for tailoring digital interventions to individual patient needs.

Purpose of the Study:

• To investigate personalization strategies for digital depression treatment using transfer learning and subgroup models.
• To evaluate the efficacy of subject-dependent standardization in predicting and forecasting depressive symptoms.
• To assess model fairness across patient subgroups in digital mental health interventions.

Main Methods:

• Utilized a longitudinal dataset from depression patients undergoing digital intervention.
• Incorporated passive mobile sensor data and ecological momentary assessments for modeling.
• Employed transfer learning, subgroup models, and subject-dependent standardization for personalized predictions.

Main Results:

• Subject-dependent standardization improved end-of-day PHQ-2 prediction by 25% (MAE $$) compared to baseline.
• Transfer learning enhanced one-day-ahead forecasting accuracy (MAE $$), improving the baseline by $$.
• Personalization strategies resulted in fairer group-level models.

Conclusions:

• Personalization via subject-dependent standardization and transfer learning significantly improves prediction and forecasting of depressive symptoms.
• Personalized machine learning architectures offer a promising avenue for enhancing digital depression interventions.
• Further research is needed to address technical and clinical limitations for broader implementation.