Experience and future perspective of artificial intelligence (AI) in postoperative gastric cancer treatment recommendation: a retrospective, observational, cross-validation study with multidisciplinary team board

Abstract

Abstract Background: Recently, owing to significant growth in the amount of information produced by cancer research, staying abreast of the developments has become a challenging task. Artificial intelligence (AI) can learn, reason, and understand the enormous corpus of literature available to the scientific community. However, large-scale cross-validation studies comparing the recommendations of AI and multidisciplinary tumor boards (MTB) in gastric cancer treatment have rarely been performed. Therefore, we retrospectively conducted a real-world study to assess the level of concordance between AI and MTB treatment recommendations. Methods: We retrospectively analyzed the treatment recommendations of Watson for Oncology (WFO) and MTB for 322 patients with gastric cancer from January 2015 to December 2018 and compared the degree of agreement between them. The patients were divided into concordance and non-concordance groups. The factors affecting the concordance rate were analyzed. Results: The concordance rate between AI and MTB was 86.96% at consideration level (280/322). The concordance rate for stage I gastric cancer was the highest (96.93 %). The concordance rates for stages II and III were 88.89% and 90.91%, respectively, which were close to 90%; however, the concordance rate for stage IV was the lowest at 45.83%. In the multivariate analysis, age, performance status, and stage IV gastric cancer had a significant effect on concordance between MTB and WFO. Conclusions: The factors affecting the concordance rate were age, performance status, and stage IV gastric cancer. For increasing the validity of future medical AI systems for gastric cancer treatment, their supplementation of the local guidelines and the ability to comprehensively understand individual patients is essential.