eP132 - Generating a Machine Learning Predictive Tool for Adverse Events Following Colorectal Cancer Resection

Purpose/Background: There is variation in the risk of adverse events following colorectal cancer surgery. Adverse events can meaningfully worsen patient prognosis and result in increased healthcare costs. To inform referral decisions and optimize outcomes following colorectal surgery, we developed a prediction model to estimate the occurrence of adverse events using a single health system's data.Methods/Interventions: We analyzed a five-hospital university inpatient claims database from 2017-2022 to build a predictive model for adverse events following colorectal cancer resection. We defined adverse events as postoperative complications not present on admission, prolonged length of stay (greater than 10 days), unplanned 30-day readmission, or death. Postoperative complications were defined using standard ICD-10 definitions. The models included 81 patient and provider features, such as tumor location, demographics, social determinants of health, clinical risk factors and surgeon volume. We used 80/20 splits to train and test the model using the Gradient Boosting machine algorithm and a ten-fold repeated cross-validation process. Finally, we used an isotonic regression to calibrate the model's probabilities.Results/Outcomes: In total, 2301 colorectal resections were included in our cohort during the study period. Patient median age was 67 years (IQR = 55-75), with 1128 (49%) female and 1790 (78%) Caucasian patients. The postoperative complication rate was 20.4%, median length of stay was 5 days (IQR 4-8 days), 30-day readmission rate was 9.9%, and mortality rate was 1.3%. In the final model, 41 features demonstrated importance (see figure). Our model achieved an AUC of 0.74 with a calibrated Brier score of 0.161 and an expected calibration error of 0.0.Conclusion/

Discussion: A novel predictive model for adverse events following colorectal cancer resection using unique features from a single academic institution’s claims data demonstrated excellent accuracy and superior calibration. This model will be used in a prototype clinical care tool designed to motivate the mitigation of modifiable risk before colorectal resection and inform referral decisions.Figure 1: Feature Importance Plot; Top 17 Important Features