Exploring the Potential Role of Upper Abdominal Peritonectomy in Advanced Ovarian Cancer Cytoreductive Surgery Using Explainable Artificial Intelligence

Abstract

The Surgical Complexity Score (SCS) has been widely used to describe the surgical effort during advanced stage epithelial ovarian cancer (EOC) cytoreduction. Referring to a variety of multi-visceral resections, it best combines the numbers with the complexity of the sub-procedures. Nevertheless, not all potential surgical procedures are described by this score. Lately, the European Society for Gynaecological Oncology (ESGO) has established standard outcome quality indicators pertinent to achieving complete cytoreduction (CC0). There is a need to define what weight all these surgical sub-procedures comprising CC0 would be given. Prospectively collected data from 560 surgically cytoreduced advanced stage EOC patients were analysed at a UK tertiary referral centre.We adapted the structured ESGO ovarian cancer report template. We employed the eXtreme Gradient Boosting (XGBoost) algorithm to model a long list of surgical sub-procedures. We applied the Shapley Additive explanations (SHAP) framework to provide global (cohort) explainability. We used Cox regression for survival analysis and constructed Kaplan-Meier curves. The XGBoost model predicted CC0 with an acceptable accuracy (area under curve [AUC] = 0.70; 95% confidence interval [CI] = 0.63-0.76). Visual quantification of the feature importance for the prediction of CC0 identified upper abdominal peritonectomy (UAP) as the most important feature, followed by regional lymphadenectomies. The UAP best correlated with bladder peritonectomy and diaphragmatic stripping (Pearson's correlations > 0.5). Clear inflection points were shown by pelvic and para-aortic lymph node dissection and ileocecal resection/right hemicolectomy, which increased the probability for CC0. When UAP was solely added to a composite model comprising of engineered features, it substantially enhanced its predictive value (AUC = 0.80, CI = 0.75-0.84). The UAP was predictive of poorer progression-free survival (HR = 1.76, CI 1.14-2.70, P: 0.01) but not overall survival (HR = 1.06, CI 0.56-1.99, P: 0.86). The SCS did not have significant survival impact. Machine Learning allows for operational feature selection by weighting the relative importance of those surgical sub-procedures that appear to be more predictive of CC0. Our study identifies UAP as the most important procedural predictor of CC0 in surgically cytoreduced advanced-stage EOC women. The classification model presented here can potentially be trained with a larger number of samples to generate a robust digital surgical reference in high output tertiary centres. The upper abdominal quadrants should be thoroughly inspected to ensure that CC0 is achievable.

Keywords: complete cytoreduction; epithelial ovarian cancer; explainable artificial intelligence; machine learning; survival; upper abdominal peritonectomy.

Figure 1

(A) Receiver Operator Characteristic (ROC) curve showing the diagnostic accuracy of all the surgical sub-procedures for the prediction of complete cytoreduction (AUC = 0.63) (B) Precision Recall curve and Average Precision performance value (AP = 0.44).

Figure 2

Model classification differences explained by the SHAP values. (A) Summary plot showing feature distribution plots based on the sum of SHAP value magnitudes over all samples. The color represents the feature value (Red not CC0 or no, Blue CC0 or yes resection) and the x-axis represents the impact score according to binary output (B) Standard bar plot of the mean absolute SHAP values for each feature showing the average impact on the global model output. SHAP, Shapley Additive explanations; CC, Complete Cytoreduction.

Figure 3

(A) Feature importance plot showing the relevance of each variable to the CC0 prediction when screened using random forest. (B) Correlation heatmap demonstrating the pairwise correlations amongst the surgical procedures. The Pearson correlation ($r^2$) was used. CC; complete cytoreduction.

Figure 4

Dependence plots demonstrating clear inflection points for several surgical sub-procedures at cytoreduction (A–C) Upper abdomen, (D–F) Bowel resections.

Figure 5

Dependence plots demonstrating clear inflection points for various regional lymph node dissections.

Figure 6

Performance metrics of devised models for the prediction of complete cytoreduction. (A) UAP. (B) Composite model comprised of UAP and commonly used engineered features.

Figure 7

Cohort survival outcomes analyzed according to the occurrence of UAP (blue = UAP cohort; orange=non-UAP cohort) (A) progression-free-survival (B) overall-survival. Note the shape difference between the concave (UAP group) and the sinusoidal (non-UAP group) curves. Hazard ratio (HR) and 95% confidence interval (CI) for prospective log-linear associations (Cox regression) between (C) recurrence and non-recurrence (D) fatal and non-fatal outcomes including the UAP and commonly used engineered features. The shape of the curves rather than the hazard ratio can be used to quantify the benefit from the intervention. In contrast, a relatively small hazard ratio (concave curves) can yield large intervention effects reflected by longer median survival times for 50% of patients.

Figure 8

Study flowchart. The probability to achieve complete cytoreduction (CC0) can be well quantified by a ML-driven model inclusive all surgical sub-procedures. Upper abdominal peritonectomy is the most important predictive feature. A “surgical package” of maximal effort targeted cytoreduction including upper abdominal peritonectomy should be offered in selected patients. Thorough inspection of upper abdominal quadrants to ensure that CC0 is achievable reflects good clinical practice. ML: Machine Learning.