Development and explainable AI-driven characterization of a prognostic model for haploidentical transplantation outcomes

Abstract

We developed a novel, explainable artificial intelligence (AI)-driven prognostic model using a contemporary single-centre cohort of 668 patients undergoing haploidentical hematopoietic cell transplantation (HCT) with post-transplantation cyclophosphamide (PTCy) prophylaxis (2015–2024) to define the hierarchy of prognostic factors for overall survival. Employing a gradient boosting machine with explainable AI techniques, including partial dependence plots and surrogate decision trees, revealed a non-linear, U-shaped effect of donor age, with optimal outcomes for donors in their late 20 s to early 40 s. Recipient age was the dominant risk factor, modulated by donor age and HLA factors. HLA-DPB1 non-permissive mismatches were the most significant HLA factor (9.6% 3-year OS reduction), followed by HLA-B leader and HLA-DQB1 mismatches, with no independent effect of HLA-DRB1. The model accurately stratified patients into distinct risk quartiles, with 3-year survival ranging from 75% in the lowest-risk to below 20% in the highest-risk group. Critically, simulation analyses demonstrated that optimal donor selection (age 30, no key HLA mismatches) could shift a patient’s risk profile by approximately one full quartile, significantly improving survival (e.g., Q3 patients from 20% to 50% 3-year OS). This study provides a practical, data-driven framework for personalized donor selection in the modern HCT era. Future multi-center validation studies are warranted.

Introduction

The widespread adoption of post-transplantation cyclophosphamide (PTCy) for graft-versus-host disease (GVHD) prophylaxis has dramatically improved outcomes in HLA-haploidentical hematopoietic cell transplantation (HCT), leading to its emergence as the most commonly utilized mismatched donor source¹. The safety and efficacy of this approach have expanded the donor pool exponentially, ensuring that nearly every patient has multiple suitable haploidentical family members. This success, however, has created a new clinical challenge: with multiple potential donors available, the optimal selection strategy remains poorly defined.

While non-HLA factors like donor age are established predictors of outcome in the setting of HLA-matched unrelated donor HCT with conventional calcineurin-inhibitor-based GVHD prophylaxis^2–8, their significance in the haploidentical-PTCy context is complicated by inconsistent findings in the literature^9–20. For instance, a single institution analysis reported that increasing donor age was associated with poorer survival⁹. Notably, this analysis was limited to patients receiving purely non-myeloablative conditioning within a historical cohort (2002–2017). Other large studies by the Center for International Blood and Marrow Transplant Research (CIBMTR) have shown a context-dependent effect, where donor age’s influence varied significantly by conditioning intensity and recipient age¹⁸, or was even negated after accounting for recipient age²⁰. While the seminal CIBMTR study by Fuchs et al.¹⁰ systematically investigated HLA factors, it was not designed to examine the joint effects of donor age and HLA mismatches.

Moreover, these studies were limited by their reliance on patients treated on or before 2017, which represents a critical gap, as this era predates the widespread use of modern therapies such as letermovir for CMV prophylaxis and ruxolitinib, belumosudil, and axatilimab for refractory GVHD, which have transformed post-HCT care. Furthermore, the complex and non-linear interplay between donor age, recipient age, and HLA factors underscores the need for advanced modeling. Therefore, an interpretable machine learning approach is best suited to determine whether the prognostic weights of these factors have shifted in the contemporary era and to translate the complex relationships into clinically actionable insights.

The primary objective of this study was to develop a prognostic model using a contemporary, single-centre cohort of patients undergoing haploidentical HCT between 2015 and 2024. By employing an explainable machine learning approach using a gradient boosting machine (GBM) model, we aimed to not only define the current hierarchy of prognostic factors but also provide a transparent, clinically interpretable tool for personalized risk stratification in the modern therapeutic landscape, addressing gaps in prior studies and enhancing donor selection strategies.

Results

A total of 668 consecutive patients undergoing haploidentical HCT were included in the analysis. The median follow-up for the entire cohort, calculated by the inverse Kaplan–Meier method, was 52.8 months (95% confidence interval: 49.7–62.0). The median recipient age was 52.0 years, and the median donor age was 34.0 years. The majority of patients had acute myeloid leukemia or myelodysplastic syndrome (59%), presented with a low/intermediate Disease Risk Index (65%), and received myeloablative conditioning (62%). While most patients had a high Karnofsky Performance Status ≥90 (65%), a significant proportion had an HCT comorbidity index score ≥3 (53%). Bone marrow (BM) was the predominant graft source (70%). Regarding HLA characteristics, GVH-directed mismatches were frequent at the HLA-DRB1 (87% of pairs) and HLA-DQB1 (76%) loci. A non-permissive HLA-DPB1 mismatch was identified in 15% of pairs, while a B Leader mismatch was present in 36%. Full baseline characteristics are detailed in (Table 1).

Table1.

Baseline patient characteristics

Characteristica	Overall	Q1 (lowest risk)	Q2	Q3	Q4 (highest risk)	p valueb	p valuec
	N = 668	n = 167	n = 167	n = 167	n = 167
Patient age, years	52 (37, 61)	39 (28, 46)	49 (33, 59)	57 (45, 64)	60 (54, 66)	<0.001	<0.001
Donor age, years	34 (25, 43)	33 (22, 43)	31 (22, 39)	34 (26, 43)	37 (28, 52)	<0.001	<0.001
Disease						<0.001	<0.001
AML/MDS	392 (59%)	83 (50%)	86 (51%)	104 (62%)	119 (71%)
ALL	99 (15%)	20 (12%)	32 (19%)	24 (14%)	23 (14%)
Others	177 (26%)	64 (38%)	49 (29%)	39 (23%)	25 (15%)
Disease risk index						<0.001	<0.001
Low/intermediate	435 (65%)	165 (99%)	126 (75%)	103 (62%)	41 (25%)
High/very high	233 (35%)	2 (1.2%)	41 (25%)	64 (38%)	126 (75%)
Conditioning intensity						<0.001	<0.001
MAC	417 (62%)	122 (73%)	114 (68%)	95 (57%)	86 (51%)
RIC/NMA	251 (38%)	45 (27%)	53 (32%)	72 (43%)	81 (49%)
Graft type						0.55	0.389
Bone marrow	467 (70%)	114 (68%)	111 (66%)	120 (72%)	122 (73%)
Peripheral blood	201 (30%)	53 (32%)	56 (34%)	47 (28%)	45 (27%)
HCT comorbidity index						<0.001	<0.001
0	122 (18%)	58 (35%)	37 (22%)	16 (9.6%)	11 (6.6%)
1–2	193 (29%)	65 (39%)	60 (36%)	37 (22%)	31 (19%)
≥3	353 (53%)	44 (26%)	70 (42%)	114 (68%)	125 (75%)
Karnofsky Performance Status						<0.001	<0.001
<90	232 (35%)	37 (22%)	50 (30%)	64 (38%)	81 (49%)
≥90	436 (65%)	130 (78%)	117 (70%)	103 (62%)	86 (51%)
Donor/recipient CMV Serostatus						0.015	0.001
Neg/Neg	86 (13%)	30 (18%)	19 (11%)	19 (11%)	18 (11%)
Neg/Pos	200 (30%)	32 (19%)	55 (33%)	56 (34%)	57 (34%)
Pos/Neg	54 (8.1%)	20 (12%)	12 (7.2%)	15 (9.0%)	7 (4.2%)
Pos/Pos	328 (49%)	85 (51%)	81 (49%)	77 (46%)	85 (51%)
Donor/recipient sex mismatch						0.042	0.021
Female–male	139 (21%)	47 (28%)	35 (21%)	29 (17%)	28 (17%)
Others	529 (79%)	120 (72%)	132 (79%)	138 (83%)	139 (83%)
Interval to HCT						0.216	0.023
<6 months	184 (28%)	56 (34%)	48 (29%)	46 (28%)	34 (20%)
6–12 months	168 (25%)	38 (23%)	38 (23%)	45 (27%)	47 (28%)
>12 months	316 (47%)	73 (44%)	81 (49%)	76 (46%)	86 (51%)
Race						0.053	0.182
White	365 (55%)	78 (47%)	93 (56%)	103 (62%)	91 (54%)
Other	303 (45%)	89 (53%)	74 (44%)	64 (38%)	76 (46%)
HLA-B Leader						0.017	0.047
Match	428 (64%)	118 (71%)	115 (69%)	95 (57%)	100 (60%)
Mismatch	240 (36%)	49 (29%)	52 (31%)	72 (43%)	67 (40%)
HLA-A PBM						0.13	0.579
GvH mismatch	413 (62%)	96 (57%)	99 (59%)	116 (69%)	102 (61%)
No GvH mismatch	255 (38%)	71 (43%)	68 (41%)	51 (31%)	65 (39%)
HLA-C PBM						0.06	0.034
GvH mismatch	389 (58%)	84 (50%)	96 (57%)	105 (63%)	104 (62%)
No GvH mismatch	279 (42%)	83 (50%)	71 (43%)	62 (37%)	63 (38%)
HLA-DRB1						0.042	0.005
GvH mismatch	580 (87%)	137 (82%)	143 (86%)	145 (87%)	155 (93%)
No GvH mismatch	88 (13%)	30 (18%)	24 (14%)	22 (13%)	12 (7.2%)
HLA-DQB1						<0.001	<0.001
GvH mismatch	505 (76%)	104 (62%)	124 (74%)	129 (77%)	148 (89%)
No GvH mismatch	163 (24%)	63 (38%)	43 (26%)	38 (23%)	19 (11%)
HLA-DPB1						<0.001	0.003
No non-permissive GvH mismatch	569 (85%)	155 (93%)	148 (89%)	131 (78%)	135 (81%)
Non-permissive GvH mismatch	99 (15%)	12 (7.2%)	19 (11%)	36 (22%)	32 (19%)
Year of transplant	2019 (2015, 2021)	2019 (2015, 2021)	2019 (2016, 2021)	2019 (2015, 2021)	2017 (2015, 2020)	0.087	0.087

^aMedian (Q1, Q3); n (%).

^bOverall p value across all four quartiles.

^cp value comparing the first and fourth quartiles.

ALL acute lymphoblastic leukemia, AML acute myeloid leukemia, CMV cytomegalovirus, HCT hematopoietic cell transplantation, HLA human leukocyte antigen, MAC myeloablative conditioning, MDS myelodysplastic syndromes, Neg negative, NMA non-myeloablative conditioning, PBM protein binding motif, Pos positive, Q1 first quartile, Q3 third quartile, RIC reduced-intensity conditioning.

Donor and recipient age

The partial dependence plot (PDP), which shows how donor age affects the risk of death, revealed a non-linear, U-shaped relationship with predicted log-hazard, diverging from the common assumption of a linear risk increase with advancing donor age (Fig. 1). The nadir was observed for donors in their late twenties to early forties, with risk escalating for those younger than their late twenties and more pronouncedly for those in their forties and older. In contrast, the PDP for recipient age demonstrated a monotonically increasing risk after approximately age 40 (Fig. S1).

Fig. 1. Partial dependence plot of donor age.

The plot displays the marginal effect of donor age on the predicted log-hazard of mortality as determined by the calibrated Gradient Boosting Machine (GBM) model. The y axis represents the change in the log-hazard from the mean prediction, averaged across all patients in the cohort. The solid dark blue line is a locally weighted scatterplot smoothing (LOESS) curve fitted to the partial dependence values, and the light blue shaded area represents the 95% confidence interval around this smoothed trend.

An interaction heatmap further elucidated the complex interplay between donor and recipient age (Fig. 2). A striking synergistic pattern emerged, with the highest predicted risk (yellow shading) concentrated in the pairing of older recipients with older donors. Conversely, a broad region of lowest risk (dark purple shading) was associated with younger recipients receiving grafts from donors aged from their mid-twenties to early forties. The steepness of the vertical color gradient compared to the more gradual horizontal one clearly highlights recipient age as the dominant factor in this interaction, with donor age exerting a secondary, context-dependent influence. This finding demonstrates that the prognostic value of donor age is not uniform; rather, it exerts a non-linear influence that is strongly modified by recipient age, challenging traditional paradigms.

Fig. 2. Interaction heatmap of donor and recipient age.

The heatmap visualizes the synergistic effect of donor and recipient age on the predicted log-hazard of mortality. The color of each cell corresponds to the predicted log-hazard (yhat) for a specific combination of donor and recipient age, averaged over the distribution of all other model features. Higher predicted risk is indicated by brighter yellow shading, while lower risk is indicated by darker purple shading. The white contour lines connect regions of equal predicted risk.

Impact of HLA mismatch status on predicted survival

Using marginal Ceteris Paribus plots (Latin “all other things being equal”) to isolate the effect of individual factors, our model identified a clear prognostic hierarchy among HLA mismatches. HLA-DPB1 non-permissive mismatch in the GvH direction was the most detrimental factor (Fig. 3a), reducing predicted 36-month survival by an absolute 9.6% (from 49.3% to 39.7%). HLA-B leader (Fig. 3b) and HLA-DQB1 (Fig. 3c) mismatches also emerged as significant predictors of inferior survival, associated with absolute survival reductions of 6.2% and 5.0%, respectively.

Fig. 3. Marginal predicted survival by HLA mismatches.

Ceteris Paribus plots from the calibrated GBM model show the average predicted overall survival over time by different HLA mismatches: a Non-permissive GvH DPB1 mismatch (blue line) vs those with no such mismatch (red line); b HLA-B leader mismatch (blue line) vs Match (red line); c DQB1 GvH mismatch (red line) vs those with no such mismatch (blue line); and d overlapping curves for DRB1 GvH Mismatch (red line) and no such mismatch (blue line).

In a striking contrast, the model attributed no independent prognostic weight to HLA-DRB1 mismatch (Fig. 3d). Recognizing the strong linkage disequilibrium between HLA-DRB1 and -DQB1, we formally tested for a masked interaction using Friedman’s H-statistic. The analysis confirmed the model’s finding was robust: an HLA-DQB1 mismatch increased the predicted log-hazard by an identical +0.20 regardless of HLA-DRB1 status, yielding an interaction statistic of 0.0. This demonstrates that the two effects were purely additive and that the prognostic signal within the HLA-DRB1/DQB1 loci was carried exclusively by HLA-DQB1. The impact of other HLA mismatches and patient-related factors is shown in Figs. S2-11.

Comparison with conventional HLA matching models

To validate the clinical utility of our PBM-based framework, we performed a head-to-head comparison against a conventional HLA matching model. This analysis evaluated how substituting PBM-based features for conventional HLA matching criteria affected the predictive performance of the GBM model within this cohort. The PBM-based model demonstrated somewhat superior discriminatory power (C-index 0.697 vs. 0.691). Crucially, the PBM framework reclassified a significant proportion of conventional mismatches as permissive (14.2% for HLA-A, 16.0% for HLA-B, and 23.2% for HLA-C), identifying a subgroup of patients with lower risk who would be effectively indistinguishable under standard criteria.

Risk stratification and clinical phenotypes

The model demonstrated excellent capacity for granular risk stratification. This was first visualized in a “fan plot” of predicted survival curves based on the calibrated risk score, which revealed a dramatic separation in outcomes—ranging from over 85% long-term survival for the lowest-risk patients to near-zero for the highest-risk (Fig. 4a). To assess which patient and donor profiles were associated with the best and worst predicted outcomes, we stratified the cohort into quartiles based on their calibrated GBM risk scores. Kaplan–Meier analysis confirmed outstanding prognostic separation between the four distinct survival groups (log-rank p < 0.0001), with 3-year overall survival ranging from ~75% in the lowest-risk quartile (Q1) to below 20% in the highest-risk quartile (Q4) (Fig. 4b).

Fig. 4. Risk stratification and survival prediction.

a Predicted Survival as a Function of Continuous Risk Score, where the ‘Fan plot’ illustrates the continuous spectrum of predicted outcomes, ranging from low risk (dark purple) to high risk (bright yellow), corresponding to a specific calibrated risk score. b Kaplan–Meier survival by predicted risk quartiles showing the observed overall survival stratified by the four risk quartiles derived from the calibrated GBM model (Q1: lowest risk, Q4: highest risk).

This stratification revealed distinct clinical phenotypes at the extremes of risk. Compared to patients in the lowest-risk quartile (Q1), those in the highest-risk quartile (Q4) were not only significantly older (median age 60 vs. 39 years) but also carried a much greater burden of adverse disease and comorbidity characteristics. This included a vastly higher prevalence of high/very high Disease Risk Index (75% vs. 1.2%) and more frequent major comorbidities (HCT-CI score ≥3: 75% vs. 26%; all p < 0.001) [Table 1].

Simplified heuristic for risk assessment

To translate these risk associations into a practical clinical heuristic, we trained a surrogate decision tree to approximate its risk score predictions (Fig. 5). The tree identified a “trifecta” of factors defining the extremes of risk. The lowest-risk profile (mean risk score = −1.09) was characterized by the combined presence of a ‘Low/Intermediate’ DRI, a recipient age of 50 years or younger, and a low comorbidity burden (HCT-CI 0–2). Conversely, the highest-risk profile (mean risk score = +1.11) consisted of patients with a ‘High/Very High’ DRI, who were also older than 49 and presented with a high comorbidity score (HCT-CI ≥ 3).

Fig. 5. Surrogate tree explaining the calibrated GBM model’s risk score.

This decision tree approximates the predictions of the calibrated GBM model by recursively partitioning the patient cohort based on key characteristics. Each terminal node displays a boxplot showing the distribution of calibrated risk scores (log-hazard) for patients falling into that profile. Paths from the root to the terminal nodes define distinct patient subgroups and their associated mean predicted risk.

Mitigation analysis: investigating the mitigating effect of optimal donor selection

Given that patient and disease factors were the primary drivers of risk, a critical question emerged: to what extent can an optimally selected donor mitigate this intrinsic risk? To investigate this, we simulated a ‘best-case’ versus ‘worst-case’ donor scenario for each patient risk quartile. Drawing on our model’s data-driven insights, we defined a ‘best-case’ donor as age 30 (the optimal nadir from our PDP) with no mismatches at the prognostically significant DPB1, DQB1, and B Leader loci. Conversely, a ‘worst-case’ donor was defined as age 50 (an age with increased predicted risk) and mismatched at all three key loci.

The simulation revealed that the impact of donor selection was highly dependent on the patient’s baseline risk, but its potential to improve outcomes was profound. Quantitatively, the analysis demonstrated that optimal donor selection could consistently shift a patient’s risk profile by approximately one full quartile. For patients in the Q2 quartile, this translated to a 3-year survival increase from ~40% to 70%. The benefit was even more pronounced for Q3 patients, whose survival more than doubled from ~20% to 50%. Finally, even for patients in the highest-risk Q4 quartile, an optimal donor could triple their chance of survival at 3 years from around 10% to 30%. These survival improvements are illustrated in the predicted survival curves (Fig. 6).

Fig. 6. Predicted impact of donor selection on survival by patient risk quartile.

This plot illustrates the simulated effect of optimal versus sub-optimal donor selection on predicted survival curves, stratified by patient risk quartile (Q2, Q3, Q4). Within each quartile, the blue line represents predicted survival with a ‘Best-Case Donor’ (age 30 years, no detrimental HLA mismatches), while the red line represents predicted survival with a ‘Worst-Case Donor’ (age 50 years, all detrimental HLA mismatches). The vertical distance between the curves within each panel quantifies the potential for risk mitigation through donor selection in that patient risk subgroup.

Simulation of risk mitigation

Finally, to dissect the specific contribution of each donor factor, we performed a one-at-a-time sensitivity analysis. This investigation revealed a clear hierarchy: the impact of avoiding key HLA mismatches was consistently dominant over optimizing donor age for mitigating risk (Fig. 7). The most striking example was the HLA-DPB1 mismatch. For intermediate-risk patients (Q2), selecting a donor without a non-permissive DPB1 mismatch yielded the greatest absolute risk reduction (Δ log-Hazard ≈ 0.34). This benefit significantly outweighed that of avoiding other individual mismatches and was substantially greater than the risk reduction from optimizing donor age alone.

Fig. 7. Impact of favorable donor selection on predicted risk.

This bar chart quantifies the absolute reduction in predicted log-hazard achieved by having a favorable donor characteristic (DQB1: no mismatch vs. mismatch; DPB1: no non-permissive vs. non-permissive mismatch; donor age: optimal 30 vs. sub-optimal 50; B Leader: match vs. mismatch). The reduction is calculated for each donor factor within patient risk quartiles Q2, Q3, and Q4, highlighting how the magnitude of benefit varies across patient risk levels.

Discussion

In this contemporary cohort, our explainable AI model established a clear prognostic hierarchy, confirming that patient- and disease-related factors remain the primary drivers of overall survival after haploidentical HCT. However, our analysis also revealed that donor characteristics exert a significant, secondary influence and represent a key opportunity for risk mitigation. Among these donor factors, favorable HLA matching was consistently more impactful than optimizing donor age. We also uncovered a non-linear, U-shaped relationship for donor age, with optimal outcomes for donors in their thirties, challenging the traditional view. Furthermore, we identified a synergistic interaction between donor and recipient age, where the risk from an older donor was most pronounced in older recipients. This complex interplay underscores that an approach to donor selection by simply dichotomizing donors as ‘younger’ vs ‘older’ is insufficient and highlights the value of our machine learning model.

The clinical implications of our findings are substantial. First, the U-shaped non-linear effect of donor age suggests that both very young and older donors may pose increased risks, potentially due to immunological immaturity (an extreme example being cord blood cells^21,22) or age-related immunosenescence^23,24 and declines in graft quality, respectively. Second, our model’s ability to stratify patients into distinct risk quartiles, with 3-year survival ranging from 75% in the lowest-risk group to below 20% in the highest-risk group, enables precise risk assessment.

More importantly, our analyses provide a clear, data-driven framework for mitigating that risk. Our sensitivity analysis reveals that for any given patient, prioritizing favorable HLA factors—particularly avoiding a non-permissive DPB1 mismatch—was a more powerful lever for risk reduction than optimizing for donor age. The cumulative benefit of this strategic approach is profound: simulation of an optimal ‘best-case’ donor (age 30 s with no key HLA mismatches) can shift a patient’s risk profile by approximately one full quartile, significantly improving survival. For an intermediate-risk patient (Q3), this translates to more than doubling their 3-year survival probability from 20% to 50%. These findings empower clinicians to move beyond simple heuristics and make truly nuanced, data-driven donor selection decisions.

Our findings regarding specific HLA mismatches both build upon and diverge from prior landmark studies, likely reflecting our contemporary cohort and distinct analytical approach. In striking contrast to the seminal CIBMTR analysis by Fuchs et al.¹⁰, which reported that non-permissive HLA-DPB1 mismatches were beneficial for survival, our model identified these same mismatches as detrimental. Similarly, our findings for the HLA-DR/DQ loci also differed. Whereas Fuchs et al.¹⁰ described a beneficial interactive effect between DRB1 mismatch and DQB1 matching on progression-free survival but not on OS, our model found no such interaction. Instead, we identified an independent, detrimental effect of HLA-DQB1 mismatch on OS, with no significant contribution from HLA-DRB1.

A potential reason for these differing observations may be attributable to the significant evolution in post-HCT care. Our contemporary cohort (2015–2024) uniquely captures an era with widespread use of modern therapies for GVHD (ruxolitinib, belumosudil, axatilimab) and CMV prophylaxis (letermovir), which were largely unavailable to the historical 2008–2017 Fuchs cohort. It is plausible that superior management of transplant complications now unmasks the intrinsic, detrimental alloreactivity driven by certain class II mismatches.

Moreover, it is important to contextualize our findings regarding HLA mismatches and donor age within the specific immunobiological setting of our cohort. Our study population predominantly received BM grafts (70%) and myeloablative conditioning (62%). This stands in contrast to the previous CIBMTR analysis¹⁰, which included a higher proportion of peripheral blood stem cell (PBSC) grafts (57%) and reduced-intensity conditioning (56%). Since PBSC grafts typically contain a higher T-cell load than BM, the clinical impact of specific HLA mismatches – particularly immunogenic Class II disparities like HLA-DPB1 – may be amplified in PBSC transplants due to increased alloreactivity.

We acknowledge our study limitations. As a single-center analysis, our findings may be influenced by institutional factors unique to a tertiary care referral center. This includes not only our specific patient demographics and transplant protocols but also our routine use of advanced therapies for post-HCT relapse prevention and access to novel clinical trials. The application of these specialized strategies might differ from ‘real-world’ community practice and could limit the generalizability. Furthermore, future multi-center studies are needed to apply this explainable AI framework to secondary outcomes (e.g., GVHD, relapse, non-relapse mortality) and to validate these findings in broader populations, and in specific subsets, particularly in cohorts with older patients or those utilizing PBPC grafts. While our model was rigorously cross-validated during hyperparameter tuning, the reported performance metrics represent internal validation. Therefore, external validation in such diverse practice settings is essential to confirm generalizability.

The use of XAI in this study represents a significant advance in HCT research, bridging the gap between complex machine learning models and clinical applicability. The surrogate decision tree, which distils the model into a simple heuristic based on recipient age, DRI, and comorbidity burden, offers a practical tool for clinicians. Looking forward, our findings pave the way for other studies to validate the model and evaluate its impact on clinical outcomes. Furthermore, integrating additional factors, such as immunologic markers or immune reconstitution, could enhance the model’s predictive power.

In conclusion, with the use of novel, explainable AI-driven prognostic model for haploidentical HCT, we demonstrate that strategic donor selection—prioritizing younger donors without key HLA mismatches—can significantly improve survival across risk profiles. These findings offer a practical framework for personalized risk stratification and clinical decision-making. Despite the single-center design, the model’s robust performance and interpretable insights lay a strong foundation for future multi-center validation studies.

Methods

Study population

We analyzed a cohort of all consecutive patients (N = 668) who underwent haploidentical HCT with PTCy prophylaxis at the MD Anderson Cancer Center between 2015 and 2024. The study was performed in accordance with the Declaration of Helsinki and approved by the local Institutional Review Board (IRB 2025-0916), which granted a waiver for informed consent due to the retrospective nature of the study. Our primary objective was to identify and quantify the relative importance of donor-related factors (e.g., donor age, HLA mismatches) versus patient-related factors (e.g., recipient age, comorbidities, disease characteristics) in influencing overall survival (OS). Patients alive at their last follow-up were censored.

Statistical analysis

All analyses were performed using R (version 4.4.2). To predict OS, we used a GBM with a Cox proportional hazards objective function. This machine learning algorithm was chosen for its ability to model complex, non-linear relationships between patient and donor factors, which traditional statistical methods like Cox regression may not fully capture. All patient, disease, donor, and transplant characteristics listed in Table 1 were included as pre-specified candidate predictors for the GBM model. No exploratory or post-hoc variables were introduced during the modeling process.

Haploidentical donors were defined as related donors with ≥2 HLA mismatches in the nonshared haplotype. HLA typing was performed for patients and donors at HLA-A, -B, -C, -DRB1, -DQB1, and -DPB1 loci using sequence-specific oligonucleotide intermediate resolution, sequence-based typing, and/or next-generation sequencing high-resolution methodologies. Class I HLA loci were categorized by peptide binding motif (PBM) status^25,26. PBM matching was performed using the CIBMTR PBM matching tool (https://pbm-matching-tool.b12x.org/). All HLA Class I mismatches were categorized by their PBM status. Specifically, class I “PBM-mismatched” group includes graft-versus-host (GVH) unidirectional mismatched or bidirectionally mismatched pairs. “Absence of PBM-mismatched” includes donor-recipient pairs matched within the same PBM group or unidirectional mismatched in the host-versus-graft (HVG) direction. B-leader mismatch was assessed using the online calculator (https://www.ebi.ac.uk/ipd/imgt/hla/matching/b_leader/) and categorized as matched or mismatched²⁷. HLA-DRB1 and -DQB1 pairs were classified based on mismatches in the GVH vector. HLA-DPB1 mismatches were classified according to TCE group utilizing the web tool (https://www.ebi.ac.uk/ipd/imgt/hla/matching/dpb_v2/), identifying mismatches as permissive, non-permissive HVG, or non-permissive GVH. Consequently, donor-recipient pairs with non-permissive GVH mismatch were labeled as “DPB1 non-permissive mismatched” while all others were categorized as “no DPB1 non-permissive mismatch”^28,29.

Data completeness was high for the majority of prognostic factors; variables such as donor age, recipient age, HCT-CI, and graft type were 100% complete. Missing data were predominantly limited to KPS (19.8%; n = 132), with negligible missingness in disease risk index (DRI, 1.3%; n = 9), donor/recipient cytomegalovirus (n = 4, 0.6%) and HLA factors: B-leader, n = 8 (1.2%) and DPB1, n = 8 (1.2%). To address this, we implemented multivariate imputation by chained equations (MICE) using the mice R package. Methodologically, imputation was performed on the entire institutional dataset of 2270 consecutive transplants (including all donor types) prior to subsetting for the haploidentical cohort. This allowed the imputation algorithm to leverage a strong correlation between variables such as HCT-CI (which was 100% complete) and KPS across a sample size more than three times larger than the study cohort, ensuring robust estimation of patient frailty. Given overall low missingness, a single imputed dataset was generated (m = 1, maxit=10). The imputation used predictive mean matching for continuous variables and polytomous logistic regression for categorical variables. Crucially, survival time and event status were excluded from the imputation model’s predictor matrix to prevent data leakage.

Model development

We developed a time-to-event prediction model using GBM, implemented via the gbm R package with a Cox proportional hazards (distribution = “coxph”) objective function.

A robust, two-stage hyperparameter optimization process was implemented. First, to efficiently search the parameter space, we employed Bayesian optimization (rBayesianOptimization package) with a 5-fold cross-validation scheme. The process identified the optimal combination of three key hyperparameters by maximizing the mean Concordance Index (C-index) across folds: interaction.depth: Search space [2,6]; Optimal value selected: 3; shrinkage (learning rate): Search space [0.001, 0.1]; Optimal value selected: 0.038; n.minobsinnode: Search space [5,20]; Optimal value selected: 20.

Second, a final GBM was configured with these optimal parameters and trained on the entire dataset. The optimal number of boosting iterations (trees) was determined using the gbm.perf function with an internal 5-fold cross-validation (method = “cv”). This process selected 83 trees for the final model. It is important to note that the performance metrics derived from this final model reflect internal validation (apparent performance) on the full development cohort, rather than validation on an independent external test set.

The raw output of the GBM is a linear predictor (risk score) that is not directly interpretable as a survival probability. To generate clinically meaningful predictions, we performed a Cox recalibration. A simple Cox proportional hazards model was fit (rms package) using the raw GBM risk score as the sole covariate. This recalibration model confirmed that the GBM score was a powerful predictor of mortality (Hazard Ratio [HR] = 2.98, 95% CI: 2.51–3.54). The final calibrated risk scores (log-hazards) and the corresponding time-dependent survival probabilities from this step were used for all subsequent analyses.

The final calibrated model’s performance was assessed using multiple metrics based on internal validation:

• Discrimination: The model’s ability to discriminate between patients with different survival outcomes was excellent, achieving an apparent (training set) Concordance Index (C-index) of 0.697. This indicates a high probability that the model correctly ranks patients by survival time.

• Overall Predictive Accuracy: We evaluated overall accuracy using the Integrated Brier Score (IBS), which measures the average prediction error over time (lower is better). Our GBM model achieved an IBS of 0.189, a substantial improvement over a non-personalised Kaplan–Meier baseline model (IBS = 0.243). The time-dependent Brier score plot (Figure A) visually confirms that the GBM model (orange line) consistently maintained a lower prediction error than the Kaplan–Meier baseline (blue line) across the entire follow-up period.

• Calibration Accuracy: Model calibration—the agreement between predicted probabilities and observed outcomes—was assessed using two methods:

• Binned Calibration: At 36 months post-transplant, patients were stratified into ten groups (deciles) based on their predicted survival. The model showed good agreement, with a Mean Absolute Calibration Error (E50) of 5.33%. At other key timepoints, the E50 was also low: 4.25% (12 mo), 5.02% (24 mo), and 3.90% (48 mo).

• Population-Level Calibration: We compared the model’s average predicted survival curve for the entire cohort against the observed Kaplan–Meier curve (Figure B). The two curves were nearly superimposable, yielding an extremely low Integrated Calibration Index (ICI) of 0.0067, a slope of 1.04, and an intercept of −0.025 (ideal values are 0, 1, and 0, respectively). This demonstrates excellent “calibration in the mean,” confirming the model accurately captures the overall survival trajectory of the population.

• A suite of explainable AI (XAI) techniques was used to interpret the “black-box” GBM model. All explainability analyses were performed on the final, Cox-recalibrated GBM model using a unified explainer object created with the survex R package. This framework allowed for consistent interpretation across different methods.

• Partial Dependence Plots (PDPs): We used the pdp and survex packages to generate PDPs, which visualize the marginal effect of key continuous predictors (e.g., donor age) on the calibrated log-hazard of mortality, averaged over the distribution of all other features.

• Interaction Analysis: The strength of two-way interactions between predictors (e.g., donor age and recipient age) was formally quantified using Friedman’s H-statistic, implemented in the iml package. This statistic measures the proportion of variance in the partial dependence of one variable that is explained by another.

• Surrogate Decision Tree: To provide a transparent, rule-based approximation of the GBM’s predictions, we trained a conditional inference tree (ctree from the partykit package) on the GBM’s calibrated risk scores. This surrogate model identifies the most important predictors and split points that differentiate high-risk from low-risk patient profiles.

• To quantify the clinical impact of donor selection, we performed a series of counterfactual simulation analyses. The patient cohort was first stratified into quartiles based on their calibrated risk score. Within each quartile, we simulated two scenarios:

• ‘Best-Case Donor’: For every patient in the cohort, we set donor-related variables to their most favorable levels (donor age 30 years, B-leader Match, No Non-Permissive DPB1 GvH mismatch, and no DQB1 GvH mismatch).

• ‘Worst-Case Donor’: We set the same variables to their unfavorable levels (donor age 50 years, B-leader Mismatch, Non-Permissive DPB1 GvH mismatch, and DQB1 GvH mismatch).

• For each scenario, we generated a new average predicted survival curve for the cohort, allowing us to quantify the potential shift in survival outcomes attributable to optimal versus sub-optimal donor selection.

• Key packages for analyses included:

• Model Training and Tuning- gbm (v2.2.2) and rBayesianOptimization (v1.2.1).

• Data Handling- mice (v3.17.0) and dplyr (v1.1.4).

• Survival Analysis and Performance- survival (v3.8-3), survcomp (v1.56.0), and survex (v1.2.0).

• Model Interpretation- rms (v7.0-0), partykit (v1.2-23), pdp (v0.8.2), and iml (v0.11.4).

• Visualization and tables- ggplot2 (v3.5.2), gtsummary (v2.1.0), ggparty (v1.0.0), and survminer (v0.5.0).

Author contributions

All authors reviewed the manuscript, provided critical feedback, and approved the manuscript. R.S.M. conceptualized and designed the study, performed the statistical analysis, interpreted data, wrote the manuscript, and had the final responsibility to submit for publication. K.C., J.C., and Y.C. provided crucial HLA data. G.R. provided crucial clinical data. Y.M.A., P.K., B.O., U.P., K.R., R.E.C., E.J.S reviewed and interpreted the data. Y.M.A., G.R., and R.S.M. had full access to the raw data.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to institutional policies regarding patient privacy and confidentiality. De-identified data may be available from the corresponding author [R.S.M.] upon reasonable request, subject to specific review and approval by the Department Chair and the Institutional Review Board (IRB).

Code availability

The complete custom computer code (R script) utilized for data processing, model training, hyperparameter tuning, and explainable AI analysis is provided within the Supplementary Appendix file.

Competing interests

R.S.M. served as a consultant reviewer for OrcaBio and received an honorarium from Lumanity for providing an expert opinion. E.J.S. serves as a consultant for Cimeio Therapeutics AG, New York Blood Centre, Celaid Therapeutics Inc, Scientific Advisor for Adaptimmune Limited, Navan Technologie, Axio Research, FibroBiologics, is on the Board of Directors/Management for National Marrow Donor Program, and License agreement for Affimed, Takeda, Rege Nexus, and is a speaker for Dava Oncology. K.R. and The University of Texas MD Anderson Cancer Center have an institutional financial conflict of interest with Takeda Pharmaceutical and Affimed GmbH. K.R. participates on the Scientific Advisory Board for Avenge Bio, Virogin Biotech, Navan Technologies, Caribou Biosciences, Bit Bio Limited, Replay Holdings, oNKo Innate, The Alliance for Cancer Gene Therapy (ACGT), Innate Pharma, and Shinobi Therapeutics. K.R. is the scientific founder of Syena.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-026-02377-z.