Novel machine learning (ML) and traditional statistical techniques in survival analysis of real-world data for allogeneic hematopoietic cell transplantation (HCT) consolidation in patients with acute lymphoblastic leukemia (ALL)

Abstract IntroductionConsolidative allogeneic hematopoietic cell transplantation (alloHCT) in acute lymphoblastic leukemia (ALL) is a potentially curative option in transplant-eligible patients (pts) achieving complete remission (CR), and offers superior outcomes in those achieving negative minimal residual disease (MRD) status. In older adults, treatment-decisions are often guided by, at least in part, the extrapolation from published data using both clinical trials and real-world outcomes. In this study, we aimed to apply machine learning (ML) based techniques in published datasets to compare their performances over the traditional statistical modeling. MethodsWe accessed publicly available CIBMTR de-identified datasets on ALL pts in first CR (CR1) with negative MRD (Wieduwilt - Blood Advances 2022, Ramanathan - BMT 2023, Abid TCT 2023). One study was excluded due to a smaller sample size, and 2 datasets were merged for the analysis. Patient outcomes were censored at 36 months. Multivariate cox regression analysis was adjusted for patient, disease, donor, graft, and conditioning covariates. For random-survival forest (RSF) and extreme-gradient boosting (XGBoost), covariates were ranked and weighted by importance. ResultsOf 1,998 pts, MAC-TBI was used in 1042 pts (52.2%), 598 (29.9%) received reduced-intensity/non-myeloablative conditioning (RIC/NMA), and remaining 358 (17.9%) received MAC-Chemo regimen. Median ages was 39 in MAC-TBI, and 60 in RIC/NMA. MAC-Chemo pts had a median age of 51 (P<0.001). Donors were 8/8 MUD in 912 pts (45%), HLA-identical sibling in 763 (38%), 140 (6.9%) haplo donors, cord(CB) in 121 (6%) and 7/8 MMUD in 84 (4.2%) of pts. Peripheral blood was used in 1575 (78%) of pts, bone marrow graft in 324 (16%) and CB in 121 (6%) pts. In traditional Cox-regression analysis of conditioning regimens, overall survival (OS) prediction was estimated at 0.599 with Harrell's C-index, and significant covariates included age > 60 years : HR 1.46 (p-value <0.001) and hematopoietic-cell transplant comorbidity index (HCT-CI) of ≥ 3+: HR 1.19 (p-value =0.048). Non-relapse mortality (NRM) prediction was estimated at 0.623. Significant covariates included age ≥ 60 (HR 1.45, p = 0.015), HCT-CI ≥ 3: HR 1.30 (p = 0.028), GVHD Prophylaxis: Reference Tacrolimus (tac)-based GVHD prophylaxis regimen, Calcinurin-inhibitor (CNI)/ methotrexate (MTX), HR 1.92 (95% CI: 1.24-2.98; p = 0.043). Relapse prediction was estimated at 0.583 and there were no statistically significant covariates. Chronic GVHD prediction was estimated at 0.581 and significant covariates included age > 60: HR 1.31 (p = 0.008), Donor-Recipient sex [M-M reference, M-F HR 0.89, F-M HR 1.21, and F-F HR 0.95] (p = 0.045). Donor-recipient CMV status: [-/- reference, -/+ 1.15, +/- 1.21, +/+ 1.09, CB+ 0.48, CB- 0.35] P <0.001),in-vivo T-cell depletion: Not used (reference), used (HR 0.74, P = 0.001), and GVHD-prophylaxis regimen, tac based reference, post-transplant cyclophosphamide (PTCy) [HR 0.56, 95% CI: 0.41-0.78], CNI/MTX: 1.70 [95% CI: 1.27-2.30], CNI/others: HR 1.50 [95% CI: 1.10-2.05] (p <-0.001). ML modeling included RSF and XGBoost. OS prediction was highest in XGBoost at 0.694. Variables ranked by importance [weights] included: Donor type [0.087], year of HCT [0.075] and conditioning regimen [0.075], recipient age in years [0.074], and HCT-CI with a weight of [0.070]. RSF scored highest for prediction of NRM estimated at 0.715, competing-risk adjusted NRM variable ranking and weights were in-vivo T-cell depletion [0.087], HCT-CI [0.069], year of HCT: [0.076], donor/recipient sex match: [0.075], recipient age in years p=0.073]. Relapse prediction was best modeled using RSF, and was estimated at 0.710. Competing-risk adjusted modeling of relapse ranked time from diagnosis to HCT [0.081], gender [0.078], HCT-CI and recipient age equally [0.069], and donor-age in years [0.063] as the most important variables. XGBoost provided the most predictive model for prediction of chronic GVHD and was estimated at 0.707. The 5 most important variables were graft type [0.115], in-vivo T-cell depletion: [0.075], GVHD prophylaxis [0.075], gender [0.061], immunophenotype [0.059]. Discussion Our study highlights superior concordance to historical methodology of outcomes analyses through explainable ML methods including RSF in NRM and relapse prediction, while XGBoost had the highest concordance in predicting OS and chronic GVHD in ALL in CR1 with MRD negative status.