All Times EDT
Keywords: machine learning, risk score
BACKGROUND: In this study, we considered creating a useful tool to assess the risk of 30-day readmission after surgery among colorectal cancer (CRC) patients using machine learning methods. METHODS: We utilized data from ACS National Surgical Quality Improvement Program (ACS NSQIP) between 2015 and 2017. Years 2015 & 2016 were training and 2017 was test set. We selected 35 past medical history (such as history of CHF, COPD, etc.) and preoperative lab value (such as BUN, WBC count, etc.) variables. CRC cases were identified using ICD codes, and surgery was identified using CPT codes. A binary variable for 30-day readmission was created. Deep learning, random forest, and gradient boosted machine learning models were run and best performing model was selected based on the AUROC. Variables were plotted according to the importance. We created a nomogram using logistic regression by including important variables with p<0.20. RESULTS: About 68105 CRC patients underwent surgery between 2015 and 2017; 53% patients were 65 years or older; and 47% were females. Majority of patients were Non-Hispanic White, had BMI = 25, and were on antihypertensive medication. History of smoking was reported by 14%, and COPD by 5.4%. There were 7183 (10.5%) patients who were readmitted within 30-days for any cause. Gradient boosted classification performed best (AUROC: 69%). Pre-op BUN, WBC, HTC, smoking hx, & pre-operative serum sodium were top predictors. CONCLUSION: We presented an interpretable model based on past medical history and pre-operative variables, which can be used to identify high risk patients so that their hospital readmission risk and financial burdens can be reduced. Interventions to prevent avoidable readmissions ought to be developed and implemented so that hospitalization costs can be lowered.