Donor Age Matters for Haploidentical HCT Patients Even After Adjusting for HLA Factors: A Machine Learning Approach

Abstract

A study by the Center for International Blood and Marrow Transplant Research (CIBMTR) assessed associations between HLA mismatching at individual loci and clinical outcomes after haploidentical donor hematopoietic cell transplantation (HCT) [1]. The HLA factors associated with superior progression-free survival (PFS) were HLA-B leader match, HLA-DRB1 mismatch, HLA-DQB1 match, and non-permissive T-cell epitope HLA-DPB1 mismatch. Although donor age was noted to be a significant predictor of overall survival, it was not identified as a risk factor for PFS; however, the study was not designed to examine the joint effects of donor age and HLA factors. An online calculator based on the PFS results is available (https://haplodonorselector.b12x.org/v1.0/) to aid in donor selection. The calculator may suggest choosing a much older donor over a younger one, for example, a 65-year-old B leader matched/DRB1 mismatched donor would be recommended over a 30-year-old HLA-B leader-mismatched/DRB1-mismatched donor.

As donor age is a known predictor of survival after haploidentical donor HCT [2-6], we hypothesized that the non-significant effect of donor age on PFS in the original study might be related to categorization, not fully capturing its impact. The complexity of haploidentical HCT, where multiple predictors (donor age and relationship) are interrelated and correlated with recipient factors (age), may make it difficult to separate the impact of donor age. Therefore, we reanalyzed the publicly available CIBMTR dataset from the Fuchs et al. study [1] using gradient boosting machines (GBM), a powerful machine learning regression methodology, with donor age as a continuous variable. As GBM models can capture complex, non-linear relationships and interactions between covariates and survival, this approach is particularly attractive. Additionally, we performed Cox proportional hazard (PH) analysis, again using donor age as a continuous variable, and compared the results of both models.

We used the publicly available CIBMTR dataset from the original publication [1]. The local Institutional Review Board (FHIRB0020181) approved the study, which was conducted in accordance with the Declaration of Helsinki. Our study population included patients with acute myeloid leukemia (AML), acute lymphoblastic leukemia, or myelodysplastic neoplasm who underwent a haploidentical donor HCT between 2008 and 2017. All patients received posttransplant cyclophosphamide (PTCy)-based graft-versus-host disease (GVHD) prophylaxis.

The primary outcome of interest was PFS. The dataset included 1434 patients. As donor age and HLA factors were key variables of interest, we excluded 1 patient with missing HLA-DRB1 data, 2 patients with missing HLA-DQB1 data, and 2 patients with missing donor age, in addition to excluding 22 patients with missing PFS data. A significant proportion of patients (54%) had missing HLA-DPB1 data. These patients were retained in the model with a separate missing category. The details of the statistical methods for the GBM and Cox PH analyses are elaborated in the supplemental file.

A GBM model, like other ensemble models, can be thought of as a black box, meaning that parameters of a model may not be directly interpretable. Rather, we rely on predictions from the GBM to explore the impact of covariates. To address this limitation while providing a user-friendly tool for utilizing the GBM model's predictive power, we developed a calculator that is available online at https://rohteshmehta.shinyapps.io/GBMBootstrap/. This calculator allows users to input patient, donor, and transplant characteristics to predict the estimated probability of PFS at 12, 24, and 36 months for each donor (Figure 1).

FIGURE 1

Open in figure viewerPowerPoint

A snapshot of the online calculator, allowing users to input specific patient, donor, and transplant characteristics to predict the estimated probability of PFS at 12, 24, and 36 months with different donor profiles. For demonstration purposes, this snapshot was edited to fit the figure dimensions. The figure displays the PFS predictions of a patient with four donors aged 30, 40, 50, and 60 years, detailing donor characteristics and the predicted PFS (with 95% confidence intervals) at 12, 24, and 36 months, along with the corresponding PFS curves.

The baseline characteristics of 1407 patients included in the study are shown in Table S1. The median patient age was 54.1 years [lower quartile (q1): 35.5, upper quartile (q3): 63.6], and the median donor age was 35.5 years (q1–q3: 27.1–45.4). The median follow-up among survivors was 34.4 months (q1–q3: 24.1–46.3).

For illustration, using the calculator, we predicted 12-, 24-, and 36-month PFS for a 65-year-old cytomegalovirus seropositive male patient with AML in first complete remission, with an HCT comorbidity index of 0, Karnofsky Performance Score > 90, and undergoing reduced-intensity conditioning haploidentical HCT with a peripheral blood graft. We noted a modest monotonic effect of donor age on PFS. Specifically, the predicted PFS at 36 months for patients with donors aged 10, 20, 30, 40, 50, 60, and 70 years was 63.6%, 59.0%, 57.7%, 55.8%, 55.8%, 50.3%, and 50.3%, respectively, with other non-HLA and HLA factors being constant across donors (HLA-B leader match, HLA-DRB1 mismatch, HLA-DQB1 match, and non-permissive HLA-DPB1 mismatch).

Next, it was noted that the impact of HLA matching appeared to outweigh the effect of donor age, especially with multiple coexistent HLA (mis)matches. For instance, the predicted PFS at 36 months was 50.3% for a recipient with a 65-year-old donor with the most favorable HLA factors (HLA-B-leader matched, -DRB1-mismatched, -DQB1-matched, and non-permissive -DP mismatch). In contrast, a recipient with a 20-year-old donor with the least favorable HLA factors (HLA-B-leader mismatched, -DRB1-matched, -DQB1-mismatched, and no non-permissive -DP mismatch) had a predicted PFS of 47.2% at 36 months. However, when only one HLA factor was unfavorable, the benefit of a younger donor age became even more evident. For example, the predicted PFS at 36 months was 57.3% with a 20-year-old donor and a single unfavorable HLA factor (HLA-B-leader mismatched), and 52.8% with a 20-year-old donor and a different single unfavorable HLA factor (HLA-DRB1-matched). These findings are noteworthy as they challenge the conventional assumption that the 20-year-old donors with certain unfavorable HLA factors would have worse PFS compared with the 65-year-old donor with highly favorable HLA factors, as donor age is not considered in that model.

The findings from our GBM model were corroborated in the Cox PH model, which showed that donor age was an independent significant predictor of PFS (Table S1). In multivariable Cox regression analysis, the hazards of experiencing an event increased by 0.08% with every year increase in donor age after adjusting for other covariates. HLA-DRB1-mismatching was associated with superior PFS [hazard ratio (HR) 0.67, 95% confidence interval (CI) 0.54–0.83, p = 0.0002)], while HLA-B-leader mismatching (HR 1.21, 95% CI 1.06–1.41, p = 0.007) was associated with inferior PFS. The effect of non-permissive DP mismatching (HR 0.77, 95% CI 0.57–1.04, p = 0.09) was somewhat less pronounced in our model than what was originally reported.

Our results provide novel insights into the influence of donor age in haploidentical donor HCT, considering other HLA and non-HLA factors. The previous analysis of haploidentical transplantation [1] focused on identifying key HLA factors associated with clinical outcome. In the current study, we more fully examined the influence of donor age together with HLA on outcome. Our results demonstrate the importance of incorporating donor age into selection criteria, and indicate that younger donors with less favorable HLA combinations can achieve comparable, if not better, outcomes, expanding donor selection possibilities.

Both GBM and Cox PH models consistently supported these findings, reinforcing their robustness. This consistency across diverse methodologies suggests that our results are not artifacts of a particular statistical approach, providing support for internal validity. Several methods comparing the two models (supplemental file) showed that the GBM outperformed the Cox PH models with better predictive abilities, suggesting its superior ability to distinguish between individuals likely to experience an event and those who will not.

As a retrospective analysis of registry data, our study is subject to biases related to data collection, donor selection, missing information, and confounding factors. Missing data for HLA-DPB1 (54%) and donor-recipient relationships (33%) were substantial and were accounted for by creating separate missing categories. Additionally, we were unable to perform external validation of our prediction model, limiting its generalizability. Final donor selection should take into account the most accurate estimate of patient outcomes combined with ethical and logistic issues, such as risk to the donor during the donation process, optimal timing of the HCT, and donor scheduling. At last, our focus on PFS as the primary outcome aligns with the original CIBMTR calculator. Future research could address these limitations by using larger datasets and validating our findings in external cohorts, examining other important outcomes such as overall survival, relapse, non-relapse mortality, and graft-versus-host disease.

In conclusion, our study underscores the complexity of simultaneously considering multiple characteristics of potential donors and the importance of considering donor age in selecting donors for haploidentical HCT. By employing GBM and Cox PH models, we revealed a nuanced relationship between donor age and HLA matching in predicting PFS. While HLA factors remain crucial, our findings suggest that younger donors with less favorable HLA matching can achieve comparable outcomes to older donors with optimal HLA (mis)matching. Our results also highlight the absence of a universal donor age cut-off and emphasize the need to assess the impact of donor age in conjunction with other HLA and non-HLA factors when making informed donor selection decisions.