Development of an artificial intelligence algorithm for adjuvant chemotherapy based on a nationwide registry of patients with gastric cancer by the Japanese Gastric Cancer Association (JGCA).

Abstract

4063 Background: Recent marked advances in machine learning have led to expectations of the clinical application of artificial intelligence (AI) to support medical care. Methods: A survival analysis model which consisted of 31 covariates was adopted for AI algorithms using machine learning and these were constructed using clinical data sets for training. The performance of AI algorithms was evaluated in order to determine the optimal chemotherapy including surgery alone without any adjuvant chemotherapy with the highest survival rate for each patient. This involved using clinical data for verification to compare survival of an AI-recommended treatment group, for which therapy recommended by AI was actually administered, with an AI-deprecated group, for which therapy recommended by AI was not administered. Results: The clinical characteristics of 23653 patients and treatment are described in the Table from 2011 to 2018 in a nationwide registry of gastric cancer patients in Japan by the Japanese Gastric Cancer Association and were made available for this study. S-1 monotherapy was used the most frequently of all adjuvant chemotherapy in this study. We used the "restricted mean overall survival time" (RMST) of all the patients as metrics. The RMST in the AI-recommended and the AI-deprecated groups were 51.4 and 47.5 months, respectively. Patient data for the verification were matched baseline characteristics by the propensity score. This model predicted effectively overall survival after gastrectomy. The RMST of over 80 years old patients were 43.9 in the AI-recommended and 38.3 months in the AI-deprecated, respectively. This AI algorithm recommended adjuvant S-1 more frequently for patients with higher age, male, American Society of Anesthesiologists – physical status 2, Eastern Cooperative Oncology Group performance status 1, pT3/pT4, pN2/pN3, more than the upper normal limit of preoperative blood urea nitrogen, macroscopic types 2/3, differentiated adenocarcinoma, Roux-en-Y reconstruction, and Clavien-Dindo classification grade II/III. The RMST for pStage IIA/IIB/IIIA/IIIB/IIIC were 55.7/53.9/50.3/45.7/42.2 months in the AI-recommended and 55.1/50.7/43.9/36.9/32.0 months in the deprecated. This AI algorithm showed a higher survival rate in pStage III patients particularly. Conclusions: The AI algorithm could readily be integrated into clinical practice to choose adequate adjuvant chemotherapy for each patient based on each patient's baseline data because it trained and verified the nationwide registry has good predictive performance. [Table: see text]