Predicting Clinical Outcomes at the Toronto General Hospital Transitional Pain Service via the Manage My Pain App:...

Area of Science:

• Digital health
• Machine learning
• Pain management

Background:

• Chronic pain affects over 25% globally, with complex biological, psychological, and social influences.
• Assessing and predicting chronic pain prognosis is challenging due to its subjective nature.
• Digital health apps like Manage My Pain (MMP) aid pain self-tracking and clinical support.

Purpose of the Study:

• To apply machine learning to MMP app data for predicting significant pain improvement.
• To utilize real-world user data from a transitional pain service for prognostic modeling.

Main Methods:

• Utilized 1-month of MMP app data from 160 patients, extracting 245 features.
• Developed a logistic regression model with recursive feature elimination to predict pain interference improvement.
• Employed 10-fold cross-validation for tuning and leave-one-out cross-validation for performance testing.

Main Results:

• The model achieved 79% accuracy and an AUC of 0.82 in predicting patient improvement.
• Demonstrated balanced class accuracies (sensitivity 0.76, specificity 0.82).
• All MMP app data, not solely clinical questionnaires, proved crucial for prediction.

Conclusions:

• Digital health app data integrated with clinical data can effectively predict chronic pain patient improvement.
• Machine learning holds significant potential for personalized treatment and improved outcomes in real-world clinical settings.