Predictive Time-to-Event Model for Major Medical Complications After Colectomy

Background: Planned (elective or time sensitive) colectomy are performed for indications including cancer, inflammatory bowel disease (IBD), and diverticulitis. After colectomy, patients are at risk of a variety of major medical complications, including pneumonia, myocardial infarction (MI), cerebral vascular event (CVA), venous thromboembolism (VTE), acute renal failure (ARF), and sepsis. However, different complications tend to happen at different times after surgery. Accurate risk prediction, not only whether a complication may occur in a patient, but also when, is crucial for patient education, monitoring, and disposition planning. While several studies have explored the incidence and binary risk prediction for major complications after surgeries, there has been scarce literature on time-to-complication prediction modeling in recent population cohort data. Objectives 1. To develop and internally validate Cox proportional hazards models to predict time-to-complication for each individual major medical complication captured in the American College of Surgeons National Surgical Quality Improvement Program (NSQIP) dataset (pneumonia, myocardial infarction (MI), cerebral vascular event (CVA), venous thromboembolism (VTE), acute renal failure (ARF), and sepsis) or adverse outcomes (mortality, readmission), that started within 30-days after elective colectomy. 2. To develop and internally validate machine learning models to predict time-to-complication for major medical complications and adverse outcomes (same as in objective 1) within 30-days after elective colectomy in NSQIP. The best machine learning model for each complication will be compared to the Cox proportional hazards model in terms of discrimination, and calibration. Methods: Investigators will conduct a time-to-event survival analysis in a retrospective cohort using NSQIP®, a prospectively-collected multicentre dataset with more than 150 clinical variables within 30 days after surgery. This dataset includes information on whether the patient was diagnosed with major complications (in- or out-of-hospital) as well as the number of postoperative days to the diagnoses of complications, as defined by a standardized criteria within the NSQIP operations manual. The general dataset will be linked with the NSQIP® Procedure Targeted Colectomy dataset, which contains additional colectomy-specific variables. The study will be reported according to the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guidelines and Guidelines for Developing and Reporting Machine Learning Predictive Models in Biomedical Research.