Despite data for a limited patient sample and brief history, the model could identify with 76% accuracy 3000 patients (~$480M revenue) currently not on target therapy but likely to be prescribed. Important model predictors were identified, making the solution more explainable and actionable for business.
The client launched a checkpoint inhibitor drug for skin cancer which soon became a market leader. To drive and sustain future growth within the narrow launch indication, the brand team wanted to identify and target emerging new patients currently on other therapies as likely candidates for the target brand.
The primary objective was to identify additional patients who would be eligible for the target brand leading to an increase in market share
Identification of treating or referring physicians and optimal locations or sites of care from among Community and Teaching Practices
Operationalize the solution in the form of weekly trigger-based alerts to enable sales force execution and non-personal promotional messaging towards the right physicians and accounts at the right time
A Patient Flow Analysis was performed to identify sources of business for the target brand in terms of patients switching between therapies.
ML Feature Selection & Modeling
An ML modeling framework was set up with 12 months of training and two months of test data window for an identified target universe of ~8k patients. The model features comprised time-dynamic behaviors (procedures, diagnoses, prescriber related) and static patient attributes (age, gender, payer, geography).
Numerous iterations were needed to arrive at an optimal mix of derived features. Multiple best-in-class ML classifiers (Random Forest, SVM, Gradient Boosting) were evaluated based on cross-validation accuracy while balancing false positives against false negatives.
Post Modeling Deep Insights
Profiling prescribers by HML Segment, Drug Class, Territory, and Specialty.
Relative importance of identified prescribing drivers including prior baseline therapy, days gap from diagnosis, diagnosis history, last physician specialty, and patient age.