Abstract: TH-PO408
A Machine Learning Model to Predict Patient Risk of Peritonitis Episodes
Session Information
- Dialysis: Peritoneal Dialysis - I
October 25, 2018 | Location: Exhibit Hall, San Diego Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 703 Dialysis: Peritoneal Dialysis
Authors
- Blanchard, Tommy C., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Willetts, Joanna, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- O'Connell, Michael Ryan, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Chaudhuri, Sheetal, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Usvyat, Len A., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Ellison, Brian Christopher, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Moran, Judith, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Herman, Melissa, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Dunphy, Susan M., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Maddux, Franklin W., Fresenius Medical Care North America, Waltham, Massachusetts, United States
Background
Peritonitis infections are one of the primary complications in the use of peritoneal dialysis. Predicting what patients are at a higher risk of peritonitis is of great interest so that cases of peritonitis can be caught early, or patients can be given additional training to prevent peritonitis infections altogether.
Methods
We analyzed data of 36,329 peritoneal dialysis patients who were treated by a large dialysis provider from 2016-2017. We had 10,522 cases of peritonitis over this period. We trained a machine learning model (XGBoost boosted tree model) to predict which patients will be diagnosed with a peritonitis infection in the next month based on patients’ history of peritonitis, symptoms noted by nurses during assessments, routine clinical laboratory values, and demographic data.
Results
Our machine learning model achieved an area under the ROC curve of 0.736. The features that were found to be most important for prediction were: whether the patient has had a previous peritoneal infection, how long the patient has been on peritoneal dialysis days since previous infection, changes in potassium levels, and albumin levels (Table 1).
Conclusion
We built a machine learning model that was able to predict which patients will have a peritonitis infection in the next month. This model can be used to allocate resources to try to catch infections early or prevent them. Future work can expand the features the model has access to in order to improve the model performance.
| Variable | Mean value for peritonitis | Mean value for uninfected |
| Vintage | 647 days | 621 days |
| Previous infection | 32% | 13% |
| Days since infection | 102 | 153 |
| Change in potassium | .052 | .028 |
| Albumin | 3.34 | 3.48 |
Funding
- Commercial Support – Fresenius Medical Care