Cox proportional hazards deep neural network identifies peripheral blood complete remission to be at least equivalent to morphologic complete remission in predicting outcomes of patients treated with azacitidine—A prospective cohort study by the AGMT

Abstract

AbstractThe current gold standard of response assessment in patients with myelodysplastic syndromes (MDS), chronic myelomonocytic leukemia (CMML), and acute myeloid leukemia (AML) is morphologic complete remission (CR) and CR with incomplete count recovery (CRi), both of which require an invasive BM evaluation. Outside of clinical trials, BM evaluations are only performed in ~50% of patients during follow‐up, pinpointing a clinical need for response endpoints that do not necessitate BM assessments. We define and validate a new response type termed “peripheral blood complete remission” (PB‐CR) that can be determined from the differential blood count and clinical parameters without necessitating a BM assessment. We compared the predictive value of PB‐CR with morphologic CR/CRi in 1441 non‐selected, consecutive patients diagnosed with MDS (n = 522; 36.2%), CMML (n = 132; 9.2%), or AML (n = 787; 54.6%), included within the Austrian Myeloid Registry (aMYELOIDr; NCT04438889). Time‐to‐event analyses were adjusted for 17 covariates remaining in the final Cox proportional hazards (CPH) model. DeepSurv, a CPH neural network model, and permutation‐based feature importance were used to validate results. 1441 patients were included. Adjusted median overall survival for patients achieving PB‐CR was 22.8 months (95%CI 18.9–26.2) versus 10.4 months (95%CI 9.7–11.2) for those who did not; HR = 0.366 (95%CI 0.303–0.441; p < .0001). Among patients achieving CR, those additionally achieving PB‐CR had a median adjusted OS of 32.6 months (95%CI 26.2–49.2) versus 21.7 months (95%CI 16.9–27.7; HR = 0.400 [95%CI 0.190–0.844; p = .0161]) for those who did not. Our deep neural network analysis‐based findings from a large, prospective cohort study indicate that BM evaluations solely for the purpose of identifying CR/CRi can be omitted.