Kidney Week

Abstract: TH-PO018

Machine Learning to Predict Unplanned Dialysis in Advanced CKD

Session Information

• AI, Digital Health, Data Science - I
  November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Augmented Intelligence, Digital Health, and Data Science

• 300 Augmented Intelligence, Digital Health, and Data Science

Authors

• Wong, Leslie P., Intermountain Healthcare, Salt Lake City, Utah, United States

Leslie P. Wong,

• Condie, Chenlee, Intermountain Healthcare, Salt Lake City, Utah, United States

Chenlee Condie,

• Durst, Mark J., Intermountain Healthcare, Salt Lake City, Utah, United States

Mark J. Durst,

• Mote, Kristin F., Intermountain Healthcare, Salt Lake City, Utah, United States

Kristin F. Mote,

• Sanchez, Jonathan, MDClone, Beer-Sheva, Israel

Jonathan Sanchez,

• Schneider, Michelle, MDClone, Beer-Sheva, Israel

Michelle Schneider,

• Calderone, Peter, MDClone, Beer-Sheva, Israel

Peter Calderone,

Group or Team Name

• Intermountain Kidney Services.

Background

Unplanned dialysis occurs when dialysis is initiated in the hospital setting and results in increased morbidity, mortality, and healthcare costs. Patients with advanced chronic kidney disease (CKD) are at higher risk for unplanned dialysis, but it is difficult to identify those who might benefit from targeted interventions. Machine learning was used to attempt to predict unplanned dialysis in a large CKD cohort.

Methods

A retrospective analysis of 15,424 patients in a large U.S. health system with stage 4 and 5 CKD between January 2018 to March 2023 was performed using the MDClone ADAMS Platform and a proprietary CKD temporal staging algorithm. Variables included age (71.4 ± 14.9 years), gender (Female, 55%), most recent eGFR (32.4 ± 16.7 ml/min/1.73m²), count of emergency department visits in two years prior (2.2 ± 3.3), count of all clinical encounters in two years prior (33.3 ± 35), BMI (30.2 ± 9.8), hypertension (77.3%), diabetes (58.6%), obstructive sleep apnea (34.1%), peripheral arterial disease (32.7%), and systolic blood pressure (134.4 ± 24.4 mmHg). XGBoost, a gradient boosting algorithm, was employed to predict unplanned dialysis events.

Results

The model's performance predicting unplanned dialysis was evaluated using accuracy, precision, recall, and F1 score metrics. The model achieved accuracy of 90.5%, precision of 55.2%, and recall of 54.7%, resulting in a F1 score of 0.55. Discrimination was high with an AUC 0.89 (Figure 1).

Conclusion

The model developed using a XGBoost machine learning algorithm demonstrated high accuracy and discriminatory power to identify stage 4 and 5 CKD patients at risk for unplanned dialysis. This predictive model has potential to help guide targeted interventions to prevent these events in the advanced CKD population.

Figure 1. Model Reciever Operating Characteristic Curve