Using machine learning to predict early readmission following esophagectomy

A, Major findings from the analysis of the National Readmission Database (NRD) dataset of 2016. Pyloroplasty/pyloromyotomy decreased the odds of early readmission, where later start of total parenteral nutrition (TPN) and upper esophageal neoplasms increased the odds of early readmission. B and C, Proposed method of how previous readmission data can be used in a machine learning (ML)-based model to find actionable interventions for patients (B) and quality improvement measures for surgical centers (hospitals) to decrease early readmission (C). When the decision is made to discharge a patient, the attributes are entered into the highly sensitive ML model. The ML model would then return a probability of early readmission and the variable importances. Example: Patient is deemed high risk for early readmission and top variables are pyloroplasty/pyloromyotomy was not done, TPN was prolonged, and patient developed acute kidney failure (AKF) and pneumonia (PNA). Then the team decides on the development of AKF and PNA. But we can check if there is high transit time at the pylorus using radiographic studies and we can also check the nutritional status of the patient. Based on testing, it may be decided to schedule early endoscopic dilation of pylorus and change the nutrition plan/discharge diet of the patient (B). In a quality improvement setting, hospital records on esophagectomy patients are evaluated by ML model. Example: top variables are patients on average, had high mortality and severity scores, TPNs were started later in the hospital course, and there was a high incidence of postoperative delirium among patients. Based on evaluation of workflow, it is decided that mortality and severity scores are not actionable. But the surgical center can decide to start patients' TPN earlier in the hospital course and enact measures to decrease postoperative delirium in post esophagectomy patients (C).