Abstract: PO2351
Development and Validation of an Algorithm to Predict Risk of 90-Day Hospitalization for Patients with CKD
Session Information
- Reassessing Race in Predicting Progression
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention
Authors
- Karpinski, Steph, Davita Clinical Research, Minneapolis, Minnesota, United States
- Sibbel, Scott, Davita Clinical Research, Minneapolis, Minnesota, United States
- Gray, Kathryn S., Davita Clinical Research, Minneapolis, Minnesota, United States
- Walker, Adam G., Davita Clinical Research, Minneapolis, Minnesota, United States
- Luo, Jiacong, Davita Clinical Research, Minneapolis, Minnesota, United States
- Colson, Carey, Davita Clinical Research, Minneapolis, Minnesota, United States
- Kindy, Justin M., DaVita Inc, Denver, Colorado, United States
- Bray, Tiffany L., DaVita Inc, Denver, Colorado, United States
- Brunelli, Steven M., Davita Clinical Research, Minneapolis, Minnesota, United States
Background
Patients with chronic kidney disease (CKD) are at higher risk of being admitted to the hospital than the general population. Hospitalizations in CKD patients are often associated with higher medical costs, increased morbidity, and increased risk of transition to end-stage kidney disease (ESKD). Nationally, there seems to be an increasing focus on the management of CKD upstream of ESKD. Identification of CKD patients at greatest risk of hospitalization may hold promise to improve clinical outcomes and judicious allocation of health care resources.
Methods
This model was developed using Medicare Part A and Part B claims from calendar years 2017-2019. Data from 50,000 unique patients diagnosed with CKD stages 3-5, no evidence of ESKD, or claims for dialysis were split into derivation (n = 40,000) and validation (n = 10,000) sets. The predicted outcome was all-cause hospital admissions, which occurred in 10.4% of patients 90 days after scoring. Overall performance of candidate models was assessed using area under the curve (AUC) of the receiver operating curve in addition to positive predictive value (PPV) and sensitivity across a variety of thresholds.
Results
The best model that we tested was a gradient boosting machine algorithm based on 399 input terms, which represented 147 unique clinical constructs. The model demonstrated good ability to discriminate (AUC = 0.73), which was stable when tested in a validation set (AUC = 0.73). The PPV in the validation set was 30.6%, 24.0%, and 21.6% at the 10%, 20%, and 30% thresholds, respectively. The sensitivity in the validation set was 28.8%, 45.3%, and 60.9% at the 10%, 20%, and 30% thresholds, respectively.
Conclusion
We developed an algorithm that uses only information derived from medical claims to identify CKD 3-5 patients at highest risk of being hospitalized in the near-term. This algorithm could be used as a decision support tool for clinical programs focusing on the management of CKD patient populations.