Kelsey McClure, MD, Stephanie Moyerman, PhD. St Joseph’s Hospital and Medical Center
Background: Laparoscopic cholecystectomy is one of the most common laparoscopic procedures performed today. Accurate prediction of the risks of cholecystectomy, whether open or laparoscopic, is important for informed surgical decision making, patient consent, and hospital planning. To date, large scale predictive models have been difficult to develop and validate due to lack of a unified, big data source. We utilize eleven years and over 200,000 individual cases in the NSQIP database to develop a machine learning model for predicting postoperative morbidity risks in cholecystectomy based on individual patient risk factors.
Methods: NSQIP data from the years 2005 – 2016 was queried for cholecystectomy by relevant CPT codes. The developed model utilized 28 preoperative factors as inputs: demographic data, comorbidities, and procedure related information (lap vs open and concurrent procedures). Total postoperative morbidity outcomes were combined from 17 independent fields logged in the NSQIP database. A Random Forest Classifier, an ensemble learning method, was trained on 80% of the data and tested on the remaining, hold-out 20% of the data. This allows for a “blind test” of predictive accuracy. Five fold cross-validation on the training set was used to tune the hyperparameters of the learning algorithm.
Results: A total of 245,436 cholecystectomy cases were identified, 219,4232 of which were performed laparoscopically. Of the total cases, 4.8% resulted in NSQIP defined postoperative morbidity – 3.2% of all laparoscopic and 18.7% of all open. The model produced extremely robust predictions for postoperative morbidity outcomes (c-statistic = 0.911, Brier score = 0.048), a more than 11% improvement over the NSQIP Universal Calculator. Within the model, the five most statistically significant predictive factors, in order, were laparoscopic vs open procedure, patient functional status, occurrence of sepsis 48 hours prior to surgery, inpatient vs outpatient procedure, and patient age.
Conclusions: By focusing on a specific and commonly performed procedure, we have developed a predictive model of postoperative risks that vastly outperforms the NSQIP calculator. The model can be used to provide empirical, informed patient risk scores for postoperative morbidity in cholecystectomy. The analysis also provides a determination of the most relevant factors for predicting patient risk in cholecystectomy.
Presented at the SAGES 2017 Annual Meeting in Houston, TX.
Abstract ID: 88108
Program Number: S008
Presentation Session: Outcomes/Quality Session
Presentation Type: Podium