The patient-adherence-ml-framework addresses one of the most significant challenges in life sciences: Patient Drop-off. As shown in our SHAP analysis , this engine identifies drivers of discontinuation with 85%+ accuracy, allowing commercial teams to intervene proactively.
Non-adherence isn't just a clinical issue; it is a massive financial drain. This framework moves beyond simple descriptive statistics to Predictive Intervention, identifying high-risk patients before they discontinue therapy.
- Predictive Drop-off Modeling: Utilizes XGBoost to forecast the probability of a patient missing their next refill or discontinuing therapy entirely.
- Explainable Insights (SHAP): Moves the "Black Box" of ML into transparent business logic. As seen in plot, we can pinpoint exactly how
digital_engagement_scoreorout_of_pocket_costimpacts individual risk. - Commercial Optimization: Provides actionable data for Patient Support Services (PSS) to prioritize high-risk segments for nurse call-center outreach or copay assistance.
- Modeling: XGBoost Classifier for robust, non-linear relationship detection.
- Interpretability: SHAP (SHapley Additive exPlanations) for feature importance.
- Key Features Analyzed:
digital_engagement_score: Correlates portal usage with adherence.out_of_pocket_cost: Measures financial toxicity.last_interaction_days: Recency of touchpoints with patient services.comorbidity_score: Clinical complexity impact on regimen compliance.
The SHAP summary plot illustrates the model's decision-making process:
- High Digital Engagement (Red on the left): Drastically lowers the probability of discontinuation.
- High Out-of-Pocket Costs (Red on the right): Significantly increases the risk of drop-off.
This framework serves as a critical downstream component of our broader Biotech Commercial Stack:
- Referral-Sense-AI: Finds the patient.
- Cgt-Precision-Patient-360: Verifies the biomarker and therapy fit.
- Patient-Adherence-ML: (This Repo) Ensures the patient stays on therapy.
- Net-Guard-GTN-Optimizer: Manages the revenue and rebates resulting from successful adherence.
# Install dependencies
pip install xgboost shap pandas matplotlib
# Run the adherence prediction pipeline
python src/predict_risk.py --input data/patient_claims.csv
Distributed under the MIT License.