Advancing 90-day mortality and anastomotic leakage predictions after oesophagectomy for cancer using explainable AI (XAI)

Abstract

Oesophagectomy for cancer of the oesophagus carries significant morbidity and mortality. Ninety-day mortality and anastomosis leakage are critical early postoperative problems traditionally analysed through logistic regression. In this study, we challenge traditional logistic regression models to predict results with new explainable AI (XAI) models. We used the Swedish National Quality Register for Oesophageal and Gastric Cancer (NREV) to perform traditional multivariable logistic regression and XAI. The 90-day mortality was 6.0%, while anastomosis leakage was present in 12.4%. The XAI models yielded an area under the curve (AUC) of 0.91 for 90-day mortality (as compared with 0.84 for logistic regression). For anastomosis leakage, the AUC was 0.84 using XAI (0.74 using logistic regression). We show that age (mortality increases sharply after 55 years) and body mass index (BMI) (lowest mortality for BMI 30 kg/m²) are important survival factors. Additionally, we show that surgery time (minimum anastomosis leakage for a surgery time of 200 min to sharply increase to a maximum at 375 min) and BMI (the lower the BMI, the less anastomosis leakage) are important factors for anastomosis leakage. The surgical understanding of anastomosis leakage and mortality after oesophagectomy is advanced by judiciously applying XAI to structured data. Our nationwide oesophagectomy data contains significant nonlinear relationships. With the help of XAI, we extract personalised knowledge, bringing oesophagus surgery one step closer to personalised medicine.