Predicting Patient Non-Compliance with Rare Disease Therapy
Background
- Game changing drug to treat rare disease with high patient value
- Lack of compliance with prescribed therapy leading to deviation from forecast despite hub case manger’s focus on out-of-supply patients
- KBQs: How do we mitigate patient non-compliance, and improve overall compliance rate?
Objective
Create risk scoring model that identifies the near-term non-compliance risk score for active patients and enables focused case manager action- Leverage machine learning to identify drivers of patient non-compliance, and predict the patient’s risk of non-compliance in the near-term
- Provide risk based recommendations/alerts to case managers that go beyond the typical ‘out-of-supply’ flags
Approach
- Defined non-compliance outcome (“Bad”)
- Machine Learning - Feature engineering with over 300 variables finalized based on business hypothesis and EDA results
- GLMNET, Random Forest & XGBTREE for predicting patient non-compliance scores
- Model evaluated based on ROC, sensitivity, and specificity
Insights & Results
- Fill history is strongest predictor of non-compliance
- Robust model predicts emerging non-compliance ~75% of the time
- ~80% of emerging non-compliant patients concentrated within top 4 risk deciles, with most patients not being “out-of-supply”
- Reallocation of case manager’s effort from out-of-supply to high/medium-risk patients resulted in 25% reduction in non-compliance, and slowed discontinuations