UTILIZATION OF MACHINE LEARNING FOR THE PREDICTION OF SURGICAL OUTCOMES IN COLORECTAL SURGERY
Arya Zarinsefat*, Lygia Stewart
Surgery, University of California San Francisco, San Francisco, CA
The ability to risk stratify surgical patients plays an important role in the care of colorectal surgery patients given the high incidence of surgical site infection (SSI) and overall complications. Outcome prediction is one of the major features of machine learning (ML) algorithms. The NSQIP database provides a unique opportunity for surgical data analysis, as its comprehensive nature provides a large amount patient data from which to test and validate various models upon. The large and comprehensive nature of NSQIP is ideal for the application of ML methods to identify patient predictors for informing post-operative outcomes.
We performed a retrospective analysis of the NSQIP data from 2014 and 2015. We selected all open and laparoscopic colorectal operations based on CPT codes, resulting in a total of 95,716 cases. Our outcomes of interest included a composite "all complications" variable, which included patients who experienced any SSI, wound disruption, pneumonia, unplanned intubation, pulmonary embolism, deep venous thrombosis, renal insufficiency, acute renal failure, urinary tract infection, stroke, cardiac arrest, myocardial infarction, unplanned return to operating room, or sepsis. We also analyzed a composite SSI variable which included any superficial incisional, deep incisional, or organ space SSI. The data was partitioned into a 75,000 patient training set to fit random forest, classification tree, and logistic regression models utilizing multiple pre and peri-operative variables. A test set of 20,716 patients (exclusive of training set patients) was used to predict outcomes from the fitted models, with receiver operating characteristic (ROC) curves and area under ROC (AUC) calculated for each model.
There were a total of 21,613 complications and 10,814 SSI's. For all complications, applying the fitted models to the test dataset revealed an AUC of 0.72, 0.70, and 0.73 for the random forest, classification tree, and logistic regression models, respectively. The ROC curves for each model and their respective AUC's are shown in Figure 1. For the all SSI's outcome, applying the fitted models to the test dataset revealed an AUC of 0.67, 0.59, and 0.69 for the random forest, classification tree, and logistic regression models, respectively. The ROC curves for each model and their respective AUC's are shown in Figure 2.
Our data show comparable predictive capabilities between ML methods and logistic regression. As ML methodology becomes more prevalent and clinicians become more comfortable with their applications, they may find greater utility in their clinical applications. Given the goal of prediction from ML, utilizing large datasets with multiple input variables may allow the future use of such algorithms for better prediction of surgical outcomes to help inform clinical decision making.
Back to 2021 Abstracts