Abstract 18568: Machine Learning Model for the Prediction of Early Discharge of Patient Underwent Transcatheter Aortic Valve Replacement Using Electronic Medical Record

Abstract

Introduction: With the increasing experience in TAVR, early discharge (within 24 to 72 h) has become a goal to accelerate patient recovery. Several studies suggest the possibility of safe, early discharges. But there is little consensus and considerable variation in early discharge protocols. Optimizing patient selection for early discharge can avoid prolonged hospital stays and save substantial costs without sacrificing patient safety. Hypothesis: The decision for early discharge after TAVR can be modeled and predicted by a machine learning model based on EMR data. Methods: Detailed demography, laboratory, medications, and comorbidities data from the EMR of all 1075 TAVR patients at Brigham and Women's Hospital from 2017-22 were included in a Random Forest machine learning model to make a binary prediction of whether the patient will be discharged within 36 hours after the procedure. We performed a 10-fold cross-validation experiment using the proposed model on the 1075-patient data. Receiver-operating characteristics (ROC) evaluated the prediction accuracy and uncertainty of the model, and clinical outcomes of patients with early discharge were evaluated by 6-month readmissions. Results: The mean age was 81± 9 years; 57% were male. The median length of stay was 51.3 hours. A total of 481 (44.7%) were discharged within 36 hours. Only 12 patients in this cohort were admitted with 24 hours. The average area under the ROC curve (AUC) of the 10-fold cross-validation was 0.83 (SD 0.04). (Figure) A total of 76 (7.1%) patients were readmitted due to cardiovascular reasons within 6 months, including 16/481 (3.3%) early-discharge patients vs. 60/594 (10.1%) non-early discharge patients (p<0.001). Conclusions: Our model accurately predicts early (36 hours) discharge after TAVR, indicating the feasibility of utilizing machine learning to develop an EMR-based patient risk stratification and early discharge protocol.