Psychotic Relapse Prediction in Schizophrenia Patients Using A Personalized Mobile Sensing-Based Supervised Deep...

Area of Science:

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

Background:

• Schizophrenia relapse prediction is crucial for timely intervention.
• Mobile sensing offers continuous behavioral data for relapse prediction.
• Deep learning models can capture complex behavioral patterns but may require personalization due to inter-individual differences.

Purpose of the Study:

• To develop and evaluate RelapsePredNet, a personalized Long Short-Term Memory (LSTM) neural network for schizophrenia relapse prediction.
• To compare RelapsePredNet with a deep learning anomaly detection model.
• To assess the potential of RelapsePredNet to complement existing models in a fusion approach.

Main Methods:

• Utilized the CrossCheck dataset with mobile sensing data from 63 schizophrenia patients monitored for up to a year.
• Developed RelapsePredNet, an LSTM model personalized using data from similar patients (based on demographics or baseline scores).
• Compared RelapsePredNet against a deep learning anomaly detection model and evaluated a fusion model with ClusterRFModel.

Main Results:

• RelapsePredNet significantly outperformed the deep learning anomaly detection model, showing a 29.4% improvement on the full test set and 38.8% on the Relapse Test Set (F2 scores of 0.21 and 0.52, respectively).
• The Social Functioning Scale (SFS) score was identified as the optimal metric for patient similarity-based personalization.
• A fusion model combining RelapsePredNet and ClusterRFModel improved the F2 score by 26.1% on the full test set, reaching 0.30.

Conclusions:

• Personalized deep learning models, like RelapsePredNet, show significant promise for improving schizophrenia relapse prediction.
• Mobile sensing combined with personalized LSTM models offers a viable approach for early detection and intervention.
• Fusion models integrating personalized deep learning with other machine learning techniques can further enhance predictive performance.