Machine learning evaluation of intensified conditioning on haematopoietic stem cell transplantation in adult acute lymphoblastic leukemia patients

Abstract

Background: The advantage of intensified myeloablative conditioning (MAC) over standard MAC has not been determined in haematopoietic stem cell transplantation (HSCT) for adult acute lymphoblastic leukemia (ALL) patients. Methods: To evaluate heterogeneous effects of intensified MAC among individuals, we analyzed the registry database of adult ALL patients between 2000 and 2021. After propensity score matching, we applied a machine-learning Bayesian causal forest algorithm to develop a prediction model of individualized treatment effect (ITE) of intensified MAC on reduction in overall mortality at 1 year after HSCT. Results: Among 2440 propensity score-matched patients, our model shows heterogeneity in the association between intensified MAC and 1-year overall mortality. Individuals in the high-benefit group (n = 1220), defined as those with ITEs greater than the median, are more likely to be younger, male, and to have higher refined Disease Risk Index (rDRI), T-cell phenotype, and grafts from related donors than those in the low-benefit group (n = 1220). The high-benefit approach (applying intensified MAC to individuals in the high-benefit group) shows the largest reduction in overall mortality at 1 year (risk difference [95% confidence interval], +5.94 percentage points [0.88 to 10.51], p = 0.011). In contrast, the high-risk approach (targeting patients with high or very high rDRI) does not achieve statistical significance (risk difference [95% confidence interval], +3.85 percentage points [−1.11 to 7.90], p = 0.063). Conclusions: These findings suggest that the high-benefit approach, targeting patients expected to benefit from intensified MAC, has the capacity to maximize HSCT effectiveness using intensified MAC.