Abstract: PO0971
Machine Learning-Driven Prediction of Peritoneal Dialysis Technique Failure
Session Information
- Home Dialysis: Disparities and Modality Choice
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 702 Dialysis: Home Dialysis and Peritoneal Dialysis
Authors
- Monaghan, Caitlin, Fresenius Medical Care, Global Medical Office, Waltham, Massachusetts, United States
- Willetts, Joanna, Fresenius Medical Care, Global Medical Office, Waltham, Massachusetts, United States
- Han, Hao, Fresenius Medical Care, Global Medical Office, Waltham, Massachusetts, United States
- Kraus, Michael A., Fresenius Kidney Care, Waltham, Massachusetts, United States
- Chatoth, Dinesh K., Fresenius Kidney Care, Waltham, Massachusetts, United States
- Usvyat, Len A., Fresenius Medical Care, Global Medical Office, Waltham, Massachusetts, United States
- Maddux, Franklin W., Fresenius Medical Care AG und Co KGaA, Bad Homburg, Germany
Background
Despite increased focus on starting and keeping dialysis patients on home therapy, Peritoneal Dialysis (PD) and Home Hemodialysis (HHD) rates are lower than desired. Two areas of opportunity are 1) keeping PD patients healthy so they can remain on PD longer and 2) transitioning PD patients to HHD when appropriate. To identify patients at risk of leaving PD in the short- (1-3 month) and long-term (3-6 month) timeframes, two machine learning (ML) models were developed. Along with risk scores, these models identify the factors driving increased risk to aid in prolonging time on PD while also allowing adequate notice to prepare for permanent access placement and HHD education.
Methods
Data were extracted for PD patients (n=53022) from 2016-2019; patients contributed one set of observations for each month they were on PD (n=823892 patient months). PD failure was defined as the first discharge from PD lasting over 30 days, and was coded as ‘1’ if the patient changed modality in the next 1-3 or 3-6 months for the short- and long-term models, respectively. All other observations were coded as ‘0,’ including censored events such as transplantation, loss to follow-up, or death. Two XGBoost ML models were trained using 80% of the dataset, with 20% used for evaluating model performance using 237 variables, derived from laboratory measurements, infection and hospitalization history, and other relevant clinical parameters.
Results
Evaluation of model performance on withheld data showed an area under the curve of 0.75 and 0.67 for the short- and long-term models, respectively. Patients were classified as High, Medium, or Low risk for each of their short- and long-term predictions. In the test dataset, 24% of high short-term risk patients dropped in the next 1-3 months, a rate almost 5 times higher than average and 12 times higher than low risk patients. For long-term predictions, 14% of high risk patients dropped in the next 3-6 months, 6% of medium risk, and 2% of low risk.
Conclusion
The two ML models showed good discrimination between patient risk categories for both short- and long-term timeframes. While further work is underway to gauge the clinical utility of these tools, these tools offer the potential to improve care of “failing” PD patients, reduce morbidity of transitions, and increase optimal starts with dialysis transitions.
Funding
- Commercial Support – Fresenius Medical Care