Abstract: PO0816
Predicting Time to Dialysis and Unplanned Dialysis Start Using Machine Learning Models
Session Information
- Dialysis Care: Epidemiology and the Patient Experience
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 701 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Shukla, Mahesh, CVS Health Corp, Woonsocket, Rhode Island, United States
- Garrett, Brendan Carter, CVS Health Corp, Woonsocket, Rhode Island, United States
- Azari, Ali, CVS Health Corp, Woonsocket, Rhode Island, United States
- Kipping, Emily, CVS Health Corp, Woonsocket, Rhode Island, United States
- Culleton, Bruce F., CVS Health Corp, Woonsocket, Rhode Island, United States
Background
Despite advances in nephrology care, a majority of patients are not well prepared for starting dialysis. This puts patients at a heightened risk of adverse outcomes such as increased hospitalization, higher health care costs and poorer quality of life. Most studies report prevalence of unplanned dialysis start between 40% and 60%. We have implemented a solution that allows the care team to combine their clinical judgement with the outputs of state-of-the-art machine learning models. These models learn patterns in historical data which lead to outcomes of interest.
Methods
We have developed and deployed a set of supervised machine learning models using gradient boosted decision trees that estimate the likelihood a patient with chronic kidney disease (CKD) requiring dialysis and having an unplanned start in the coming 18 months. Unplanned Dialysis Start (UDS) Model sits downstream of Time to Dialysis or Temporal Risk (TR) Model and scores the CKD patients who are predicted to need dialysis. We trained these models in the medical and pharmacy claims and lab data of 751,242 CKD patients spanning multiple years. Input features included demographics, medical history, social determinants of health, and medication adherence. We are using the model output for selection of beneficiaries in a kidney care management program. In addition, the care team is using the risk scores at the point of care.
Results
TR Model has AUC of 93% and F1-score of 0.31 whereas UDS Model has AUC of 71% and F1-score of 0.30. The models are relying on clinically relevant features in making their predictions. Top predictors include serum creatinine, serum albumin, serum phosphate, hemoglobin, CKD Stage, age, comorbidities, nephrologist visits, social determinants of health, and uremic symptoms. We are able to discover patients who are not receiving nephrology care but are at risk for an unplanned start.
Conclusion
Machine learning models developed in large claims and lab datasets can predict time to dialysis and risk of unplanned dialysis starts. These models can be integrated into care management programs to target high risk patients with interventions calibrated to the individual patient's risk. An evaluation study is the next logical step.
Funding
- Commercial Support –