Business Objective
Our client, a major provider of health insurance plans in the US, wanted to prevent avoidable hospitalizations as they carry a significant clinical and economic burden. Early intervention and effective primary care through predictive analytics assist in driving better care and disease management programs, thus preventing hospitalizations. More specifically, the client wanted to:
- Drive improvements in the existing set of ROX (risk of event) models — Risk of hospitalization and ER visits in the next 6 months, and risk of high-cost claims in the next 12 months
Challenges
- Retaining model interpretability while improving performance
- Extracting time related features for critical data elements like diagnoses and procedures
Solution Methodology
- Started with gaining better interpretation of CPT/ICD codes from
- American Medical Association (AMA): for ICD-10-PM, ICD-10-PCS, CPT codes
- Healthcare Cost and Utilization Project (HCUP): for HHS-HCC diagnosis groups, Elixhauser Comorbidity index, and CCI mapping
- Created new features to improve prediction including presence and utilization of specific diagnoses and procedures, time scale-based velocity variables, recency and multiplicity of diagnoses/procedures
- Improved model performance with advanced boosting and ML techniques – GBT classifier, isotonic regression, hyperparameter tuning
- Retained model explainability by developing interpretability modules using SHAP (SHapley Additive exPlanations)
Business Impact
- Improved models delivered 2x precision over existing models and helped identify additional USD 1M in savings in 6 months
