Skip to main content
SSO 2024 Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Presentation Icons
Recorded for On-Demand Viewing
Recorded for On-Demand Viewing
Back

Peritoneal Surface Malignancies

E338: An Explainable Machine Learning Framework for Predicting Textbook Outcome After Cytoreductive Surgery with Hyperthermic Intraperitoneal Chemotherapy

Introduction: Cytoreductive surgery with hyperthermic intraperitoneal chemotherapy (CRS-HIPEC) is the preferred treatment for select patients with peritoneal surface malignancies but has a high complication profile. Whereas traditional quality metrics in surgery are often specific and do not reflect a holistic assessment of patient outcomes, textbook outcome (TO) captures multiple key postoperative outcomes and offers a comprehensive measure of quality. As failure to achieve TO has been shown to be independently associated with worse overall survival, we aimed to develop a machine learning framework to predict TO after CRS-HIPEC.

Methods: Adult patients who underwent CRS-HIPEC were included from a prospectively maintained dataset (2001-2020). Based on consensus among surgeons from 12 academic institutions, TO was defined as the absence of: completeness of cytoreduction (CC) score >1, postoperative grade ≥2 complication, 90-day reoperation, hospital stay >14 days, non-home discharge, 90-day hospital readmission, and 90-day all-cause mortality. We trained three machine learning models (neural network, random forest, and XGBoost) and a traditional logistic regression model. The recursive feature elimination algorithm optimized the number of input variables, and SHAP values were used to assess variable importance.

Results: From the 1,954 cases analyzed, most patients were white (92%) and female (52%), with a median age of 56 years (IQR 47-64). A total of 15.8% achieved TO. The recursive feature elimination algorithm determined that optimal performance was achieved with 15 variables. After training the models using 80% of the data, random forest had the highest area under the ROC curve (AUC 0.76) in the test dataset, followed by neural network (AUC 0.75), logistic regression (AUC 0.74), and XGBoost (AUC 0.73) (Figure 1A). SHAP value analysis revealed that the most important features for predicting TO were a lower peritoneal carcinomatosis index (PCI), no major resections including proctectomy, splenectomy, partial colectomy, or small bowel resection, higher preoperative albumin levels, and absence of preoperative symptoms from peritoneal disease (Figure 1B).

Conclusions: Prediction of TO after CRS-HIPEC may be augmented with the use of machine learning frameworks. This can facilitate informed consent and help set realistic expectations. Additionally, given that postoperative complications have been linked to adverse oncologic outcomes, our model can help identify patients who may derive the most benefit from CRS-HIPEC.

Learning Objectives:

  • List the most important features for predicting textbook outcome after CRS-HIPEC.
  • List the criteria that define textbook outcome after CRS-HIPEC.
  • Describe some potential benefits of accurately predicting textbook outcome after CRS-HIPEC.