Using machine learning to predict early readmission following esophagectomy
© 2020 The American Association for Thoracic Surgery Objective: To establish a machine learning (ML)-based prediction model for readmission within 30 days (early readmission or early readmission) of patients based on their profile at index hospitalization for esophagectomy. Methods: Using the National Readmission Database, 383 patients requiring early readmission out of a total of 2037 esophagectomy patients alive at discharge in 2016 were identified. Early readmission risk factors were identified using standard statistics and after the application of ML methodology, the models were interpreted. Results: Early readmission after esophagectomy connoted an increased severity score and risk of mortality. Chronic obstructive pulmonary disease and malnutrition as well as postoperative prolonged intubation, pneumonia, acute kidney failure, and length of stay were identified as factors most contributing to increased odds of early readmission. The reasons for early readmission were more likely to be cardiopulmonary complications, anastomotic leak, and sepsis/infection. Patients with upper esophageal neoplasms had significantly higher early readmission and patients who received pyloroplasty/pyloromyotomy had significantly lower early readmission. Two ML models to predict early readmission were generated: 1 with 90.4% sensitivity for clinical decision making and the other with 92.8% accuracy and 99.3% specificity for quality review. Conclusions: We identified risk factors for early readmission after esophagectomy and introduced ML-based techniques to predict early readmission in 2 different settings: clinical decision making and quality review. ML techniques can be utilized to provide targeted support and standardize quality measures.