Predicting the Risk of Unplanned Hospital Encounters in Hemodialysis Patients: A Comparison Between Clinicians and a Machine Learning Prediction Algorithm.

Abstract

Dialysis patients have a higher risk of hospitalization than the age-matched general population. Unplanned hospital encounters (UHEs), including emergency department visits and hospitalizations, may be preventable but are difficult to anticipate. We aimed to stratify dialysis patients' risk of UHE, using a machine learning (ML)-based prediction model. We created an ML-based prediction model to stratify hemodialysis (HD) patients' risk of UHE in the subsequent 7 and 30 days. We compared the performance of this prediction model with that of clinicians. 627 patients who received maintenance HD at an academic hospital from October 2015 to April 2022 for model development and 285 patients for a comparison with clinician predictions. Demographic and clinical data as well as data derived from the dialysis treatment sessions, updated weekly. Clinicians demonstrated a higher positive predictive value (PPV) for a UHE in the subsequent 7 days (0.14 for clinicians vs 0.07 for the ML model; <i>P</i> &lt; 0.001) but lower sensitivity (0.25 for clinicians vs 0.63 for the ML model; <i>P</i> &lt; 0.001). The PPV for the 30-day UHE was comparable between clinicians and the model (0.25 for clinicians vs 0.21 for the ML model; <i>P</i> = 0.54), but sensitivity was nearly three-fold higher for our ML model relative to clinicians (0.24 for clinicians vs 0.73 for the ML model; <i>P</i> &lt; 0.001). Our ML-based prediction model was only developed within a cohort comprising HD recipients at a single center, limiting external validation. Using ML, we were able to anticipate UHE with greater sensitivity than clinician predictions with comparable PPV. Further research will determine if integration of our prediction tool into clinical care will lead to fewer UHEs. Many patients with kidney failure who receive maintenance dialysis develop illnesses that require a visit to the emergency department or admission to hospital. These unexpected events are stressful for patients and consume a great deal of health care resources. This prompted us to ask whether we could anticipate and perhaps prevent some of these hospitalizations. We developed a machine learning tool that predicted the risk of an unplanned emergency room visit or hospitalization. Using data from the dialysis treatment and general medical record, we designed a running model that predicted the risk of an unplanned hospital visit. Although imperfect, the model performed better than clinicians in predicting unplanned hospital visits.