Predicting Risk Adjustable Chronic Kidney Disease in Medicare Advantage Members
No Thumbnail Available
Authors
Boggess, Noah
Issue Date
2025-11
Type
Dissertation
Language
en
Keywords
Business, Engineering, Science, & Technological Innovation , Medicare Advantage , Chronic Kidney Disease , Machine Learning , Risk Adjustment
Alternative Title
Abstract
Chronic kidney disease (CKD) is a highly prevalent condition that contributes substantially to patient morbidity, healthcare utilization, and financial risk in value-based reimbursement models. In Medicare Advantage (MA), accurate documentation of CKD is essential both for guiding appropriate clinical management and for ensuring risk-adjusted payments that reflect population health needs. Yet CKD is often under-documented, leading to gaps in care and misaligned reimbursement. This study investigated whether predictive models trained on routinely collected administrative claims and laboratory data could identify likely cases of undocumented but risk-adjustable CKD. A quantitative, non-experimental design was employed using a dataset of MA beneficiaries from 2021 to 2023. Two modeling approaches were developed and compared: multivariate logistic regression and Extreme Gradient Boosting (XGBoost). Data preprocessing included feature selection through Lasso, Ridge, and Elastic Net regularization. Model performance was assessed on a temporally separated 2023 holdout sample to simulate real-world deployment. Evaluation metrics included accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (ROC AUC). SHAP (SHapley Additive exPlanations) values were used to interpret model predictions. Both models achieved strong predictive validity. Logistic regression yielded an accuracy of 0.896, precision of 0.855, recall of 0.771, F1 score of 0.811, and ROC AUC of 0.952. XGBoost slightly outperformed it, with accuracy of 0.900, precision of 0.842, recall of 0.806, F1 score of 0.824, and ROC AUC of 0.955. SHAP analysis confirmed that clinically meaningful predictors—particularly estimated glomerular filtration rate (eGFR), creatinine, age, and diabetes—were the dominant drivers of prediction, consistent with established knowledge of CKD progression. The findings demonstrate that interpretable predictive models can effectively surface likely undocumented CKD in MA populations, offering a scalable pathway for improving documentation accuracy, enhancing reimbursement integrity, and supporting earlier clinical intervention. This work contributes to the literature by framing predictive modeling not only as a tool for disease prediction, but also as a mechanism for addressing documentation gaps—an often-overlooked dimension of healthcare quality and equity.