Preoperative Risk Stratification in Esophageal Cancer Surgery: Comparing Risk Models with the Clinical Judgment of the Surgeon

Background: Numerous prediction models estimating the risk of complications after esophagectomy exist but are rarely used in practice. The aim of this study was to compare the clinical judgment of surgeons using these prediction models. Methods: Patients with resectable esophageal cancer who underwent an esophagectomy were included in this prospective study. Prediction models for postoperative complications after esophagectomy were selected by a systematic literature search. Clinical judgment was given by three surgeons, indicating their estimated risk for postoperative complications in percentage categories. The best performing prediction model was compared with the judgment of the surgeons, using the net reclassification improvement (NRI), category-free NRI (cfNRI), and integrated discrimination improvement (IDI) indexes. Results: Overall, 159 patients were included between March 2019 and July 2021, of whom 88 patients (55%) developed a complication. The best performing prediction model showed an area under the receiver operating characteristic curve (AUC) of 0.56. The three surgeons had an AUC of 0.53, 0.55, and 0.59, respectively, and all surgeons showed negative percentages of cfNRIevents and IDIevents, and positive percentages of cfNRInonevents and IDIevents. This indicates that in the group of patients with postoperative complications, the prediction model performed better, whereas in the group of patients without postoperative complications, the surgeons performed better. NRIoverall was 18% for one surgeon, while the remainder of the NRIoverall, cfNRIoverall and IDIoverall scores showed small differences between surgeons and the prediction models. Conclusion: Prediction models tend to overestimate the risk of any complication, whereas surgeons tend to underestimate this risk. Overall, surgeons’ estimations differ between surgeons and vary between similar to slightly better than the prediction models.