T. J. Shaik1, X. Yang2, N. Wolfrath1, G. SenthilKumar1, J. Merrill1, A. Istl1, C. Clarke1, U. N. Maduekwe1, A. N. Kothari1,2 1Medical College Of Wisconsin, Division Of Surgical Oncology, Department Of Surgery, Milwaukee, WI, USA 2Medical College Of Wisconsin, Clinical And Translational Science Institute Of Southeastern Wisconsin, Milwaukee, WI, USA
Introduction: Nearly 20% of patients require hospital readmission after undergoing cytoreductive surgery (CRS) and hyperthermic intraperitoneal chemotherapy (HIPEC). Accurate prediction of readmission could facilitate development of individualized postoperative care pathways aimed at reducing readmission risk. The primary objective of this study was to develop a machine learning model for predicting readmission risk in patients who have undergone CRS/HIPEC and integrate the final fitted model into a web-based application for clinical utilization.
Methods: Patients with peritoneal metastasis undergoing CRS/HIPEC were retrospectively identified using an institutional registry. A total of 86 features were included from the following domains: preoperative characteristics, surgical course, and postoperative complications. The final feature space was dimensionally reduced using a series of steps including model-based selection and human-in-loop feedback. Four models were trained using various algorithms including XGBoost, Random Forest (RF), Logistic Regression (LR), and Neural Network (NN) with 80% of the dataset (20% held out for validation). Model performance was assessed using predictions from the final fitted model on validation data (accuracy, AUC, precision, recall). A web-based application was developed using R Shiny.
Results: Of 322 patients, 54 were readmitted after CRS/HIPEC. Six variables were identified following feature selection: histologic grade, preoperative symptoms, low anterior resection, colon ileus, superficial surgical site infection, age. Utilizing these features, RF model performed best (accuracy 0.83; AUC .85; precision 0.98; recall 0.82) for predicting readmission. Predicted readmission probabilities were used to stratify individual-level readmission risk, with <25th quartile labeled ‘low risk’, 25th-75th quartile labeled ‘moderate risk’, and >75th quartile labeled ‘high risk’. There were 65 patients in the validation cohort, and the RF model categorized 16 patients as low risk, 32 as moderate risk, and 17 as high risk for readmission. Of those categorized as moderate and high risk, 9% and 47% were readmitted respectively; none of the low-risk patients were readmitted (Figure).
Conclusion: The RF ML model, utilizing just 6 input features, can accurately predict patients that are at risk for readmission. Using a web-based application, the model can be operationalized to provide real-time predictions. Modifying post-discharge care based on predicted readmission risk provides an opportunity to personalize care and potentially lower readmission rates in this vulnerable population.
