Passive Sensor Data Based Future Mood, Health, and Stress Prediction: User Adaptation Using Deep Learning.

Area of Science:

• Digital Health
• Machine Learning
• Wearable Technology

Background:

• Predictive wellbeing models traditionally rely on participant-specific data.
• Real-world applications require models adaptable to new users for immediate and accurate wellbeing predictions.
• Passive sensing from mobile devices and wearables offers a continuous data stream for wellbeing analysis.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting new users' mood, health, and stress using passively sensed data.
• To compare the performance of deep Long Short-Term Memory (LSTM) networks against hybrid Convolutional Neural Network (CNN)-LSTM models for wellbeing prediction.
• To assess the efficacy of transfer learning in enhancing prediction accuracy for new users, particularly with limited data.

Main Methods:

• Utilized passively collected data from wearable sensors, mobile phones, and a weather API.
• Implemented and compared deep LSTM and CNN-LSTM deep learning architectures.
• Applied a fine-tuning transfer learning approach to the deep LSTM model for new user adaptation.

Main Results:

• The deep LSTM model achieved Mean Absolute Errors (MAE) of 15.7 (mood), 15.6 (health), and 16.8 (stress) out of 100 for new users.
• Transfer learning significantly improved prediction accuracy, reducing MAE to 13.5 (mood), 13.2 (health), and 14.4 (stress).
• The transfer learning model demonstrated particular effectiveness when limited data from new participants was available.

Conclusions:

• Deep learning models, specifically deep LSTM, can effectively predict new users' wellbeing using passively sensed data.
• Transfer learning is a viable and effective strategy for adapting wellbeing prediction models to new users, enhancing accuracy and reducing data requirements.
• This approach holds promise for proactive and personalized digital health interventions.