Finding Emerging New CKD Patients
Background
- Lackluster launch performance for brand’s CKD indication
- Targeting based on prescription volume does not capture the full dynamics of the specialists who are testing and diagnosing CKD patients
- KBQs: Who are the patients at the cusp of being diagnosed with CKD? Of the doctors treating these emerging CKD patients, which ones are most likely to diagnose these CKD patients?
Objective
Build patient-level model that identifies the patients to be diagnosed with CKD in the near-term? Stratify the doctors treating these emerging CKD patients into segments based on the level of CKD testing.- Leverage machine learning to identify features of patients recent diagnosed with CKD
- Leverage features to predict the likelihood of eligible patients being diagnosed with CKD in the next six months
- Segment treating doctors into H/M/L segments based on their recent volume of CKD testing (uACR and eGFR)
Approach
- Defined CKD patient (“Good”)
- Leveraged LAAD data to define demographic, comorbidity, procedure and prescription variables
- Machine Learning – Feature engineering with over 150 variables finalized based on business hypothesis, USRDA risk factors, and EDA results
- LSTM, ANN & XGBTREE leveraged for predicting emerging CKD patients
- Model evaluated based on ROC, sensitivity, and specificity
Insights & Results
- 18 of the 150+ variables are strong predictors of emerging CKD patients
- Identified nearly 500K emerging CKD patients (‘High’ likelihood) being treated by about 10K doctors, with ‘High’ testing rates
- 90% of these 10K doctors were PCPs/ENDOs/CARDs, and only 10% are NEPHRs; fewer than 10% are sales force targets
- Expanded sales fore targets by adding non-target ENDOs/CARDS/NEPHRs within the same practice as an existing target
- Focused NPP/digital promotions on these 10K High-High doctors in addition to the High-Meduim and Medium-High doctors