A biomarker risk assessment tool predicts infection-related mortality after allogeneic stem cell transplantation

Key Points

• •MAPΔ in early allogeneic stem cell transplant independently predicts 6-month mortality because of infection.
• •We propose a 6-month infection-related mortality risk assessment tool based on MAPΔ, which guides interventions and clinical trial design.

Visual Abstract

Abstract

Nonrelapse mortality (NRM) in allogeneic haemopoietic stem cell transplantation (alloHSCT) is driven by acute graft-versus-host disease (aGVHD) and infection. The MAGIC algorithm probability (MAP) composite biomarker score is based on serum ST2 and REG3a levels at day (D)7 after transplantation and predicts aGVHD-related mortality. Although changes in MAP score predict aGVHD treatment response, their posttransplant predictive value and association with infection-related NRM (iNRM), independent of GVHD, are unclear. We evaluated the association of D0 to D21 MAP change (MAPΔ) vs D7 MAP with 6-month NRM (6mNRM) and developed a risk stratification model using random forest analysis and recursive partitioning. We prospectively enrolled 101 adult alloHSCT recipients from 2022 to 2024. Serum was collected weekly from D0 to D21 to derive MAPΔ. Associations with 6mNRM were assessed using generalized linear and random forest modeling. 6mNRM was 14.8%, with 80% (12/15) of deaths attributed to infection without prior aGVHD. Both MAPΔ ≥0.055 and D7 MAP ≥0.16 were significantly associated with 6mNRM (P < .0001), with MAPΔ showing superior predictive accuracy (receiver operator characteristic area under the curve 0.934 vs 0.779). In multivariable analysis, MAPΔ (odds ratio, 45; 95% confidence interval, 10.6-318; P = .01), but not D7 MAP, was an independent predictor of 6mNRM. A classification tree incorporating MAPΔ, D7 MAP, and age-adjusted comorbidity index stratified patients into low, intermediate, and high-risk groups. MAPΔ was a powerful, independent predictor of 6mNRM, which was driven by infection rather than aGVHD in this cohort. We propose a tool based on MAPΔ to assess iNRM, which may guide interventions and clinical trial design. Further validation in larger cohorts is warranted.

Introduction

Allogeneic haemopoietic stem cell transplantation (alloHSCT) is a powerful immunotherapeutic and curative approach for high-risk hematological malignancies. Although the long-term remission potential offers survival advantage, it is tempered by transplant-related mortality. Over the last decade, nonrelapse mortality (NRM) has remained 10% to 24% within 4 years after transplantation. Infection, graft-versus-host disease (GVHD), and conditioning-related organ damage may precipitate events that create a complex balance between treatment and inadvertent introduction of further complications, which could lead to mortality. Preventive strategies are key, but broad approaches pose challenges. For instance, generalized strategies to mitigate GVHD, for example posttransplant cyclophosphamide, are counterbalanced by an increased risk of infection and infection-related mortality.1, 2, 3 This highlights the need for individualized risk assessment to fine-tune preventive strategies. Biomarker-based and clinical predictive algorithms can assess mortality risk, allowing early detection and timely intervention to reduce mortality.

The serum proteins stimulation-2 (ST2) and regenerating islet-derived 3 alpha (REG3a) indicate inflammation, endothelial injury and organ parenchymal damage. However, neither is a consistently accurate standalone marker of acute GVHD (aGVHD) or NRM.^4,5 The MAGIC algorithm probability (MAP) score is a composite measure of ST2 and REG3a designed to overcome the limitations of each. A high MAP score, calculated on day (D)7 following alloHSCT, was shown to predict severe aGVHD-related NRM. Subsequent studies identified correlations between elevated MAP scores and non-GVHD–related complications, such as thrombotic microangiopathy.6, 7, 8 Recent studies link elevated levels of ST2 and REG3a to fatal infections, indicating a role for the MAP score in predicting non-GVHD–related infectious NRM (iNRM) after alloHSCT.^5,9

Studies assessing ST2, REG3a, and MAP score utility early after transplant have primarily assessed single time points in all transplant recipients collectively. However, evidence suggests baseline biomarker levels are influenced by patient-specific factors, including conditioning intensity, which may affect early posttransplant concentrations and MAP scores.^4,10 We hypothesize that biomarker concentrations have marked baseline heterogeneity unaccounted for in previous studies, and that measuring biomarker kinetics improves NRM prediction over single time point evaluations. In this study, our objective was to compare the association between MAP score kinetics and single time point MAP score with 6-month overall NRM (6mNRM) and iNRM. Additionally, we aim to develop a risk prediction tool for NRM that integrates clinical patient characteristics with MAP score kinetics.

Methods

Study design and population

We prospectively recruited adults undergoing consecutive elective allogeneic bone marrow transplantation at the Alfred Hospital, Melbourne, Australia, between January 2022 and January 2024. All recipients received T-cell–replete grafts. Serum samples were collected at D0 (before transplant), D7, D14, and D21 after transplant. Event-free participants were followed for at least 1 year to monitor for aGVHD, relapse, infection, and all-cause mortality.

Infection prophylaxis protocols

All alloHSCT recipients received prophylaxis against viral, fungal, pneumococcal, and Pneumocystis jirovecii infections following transplantation. Cytomegalovirus (CMV) seronegative patients received herpesvirus prophylaxis with valaciclovir until 12 months after transplant. Seronegative patients who received stem cells from a seropositive donor received weekly CMV monitoring. All CMV-seropositive patients received daily valganciclovir after neutrophil engraftment, continued until D100 after transplant, then switched to herpesvirus prophylaxis. Voriconazole was used for antifungal prophylaxis from D5 to D100 after transplant. Posaconazole was given in cases of grade 2 to 4 aGVHD or extensive chronic GVHD. Pneumococcal prophylaxis with phenoxymethylpenicillin or roxithromycin was started at inpatient discharge and continued for 6 months or until cessation of immunosuppression. Trimethoprim-sulfamethoxazole was initiated after neutrophil and platelet engraftment for Pneumocystis jirovecii prophylaxis and continued until 12 months after transplant or cessation of immunosuppression. If trimethoprim-sulfamethoxazole was not tolerated, alternatives included dapsone, atovaquone, or nebulized pentamidine. AlloHSCT recipients did not routinely receive fluoroquinolone prophylaxis.

Protein measurement and MAP score calculation

Patient serum samples were collected in serum-separating tubes, processed per the manufacturer’s instructions, and stored at −80°C within 24 hours of collection. Serum ST2 and REG3α concentrations were batch-measured using enzyme-linked immunosorbent assays (ELISA), following the manufacturer’s protocols. ST2 levels were analyzed using the human ST2/IL-33R Quantikine ELISA Kit (catalog no. RDSDST200), REG3α levels were measured with the human REG3A DuoSet ELISA (catalog no. RDSDY594005) and the DuoSet ELISA Ancillary Reagent Kit 2 (catalog no. RDSDY008B). ELISA kits were supplied by R&D Systems (Noble Park North, Australia). The composite MAP score was calculated using the published formula: log[−log(1−MAP)] = −11.263 + 1.844(log10ST2) + 0.577(log10REG3α).⁶ Changes in serum ST2 (ST2Δ), REG3α (Reg3aΔ), and MAP scores (MAPΔ) were defined as the absolute difference in concentration or score between D0 (before transplant) and D21. If a D21 sample was unavailable in the event of failed collection or patient death, the latest serum sample collected was used.

Outcomes

We employed 2 outcome measures: (1) 6mNRM as a binary outcome measure and (2) the time-to-event end point of NRM (tteNRM) as a continuous outcome measure throughout the follow-up period. NRM was defined as death from any cause that was not preceded by underlying disease relapse. Time-to-event outcomes were measured from date of transplantation to first event or date of last patient follow-up. Causes of death were collected from death certificates and review of patient medical records performed by an experienced hematologist. iNRM was defined according to the published Center for International Blood and Marrow Transplant Research definitions by Hahn et al, as occurring in patients with either no GVHD or completely resolved GVHD.¹¹ Patients were classified as having iNRM if they met the following criteria: (1) infection or infection-related organ failure was listed as the primary cause of death, (2) no active symptoms of GVHD were present, and (3) the patient was not receiving immunosuppression for GVHD at the time of death.

Statistical analyses

ST2 and REG3a concentration levels at baseline and follow-up time points were compared using 2-tailed paired t tests. The log-rank test was employed to assess differences in tteNRM based on the D7 MAP score, ST2Δ, REG3aΔ, and MAPΔ. When considering the population at risk of tteNRM, cases are censored at relapse and, in relapse-free patients not known to have died, the last date of contact. Supplementary competing risk analyses were conducted using Gray’s test. A receiver operating characteristic (ROC) curve was used to determine optimal threshold and compare discriminatory power of different tests for 6mNRM (area under the curve [AUC]). To evaluate significance of associations between MAPΔ and other transplant variables with 6mNRM, we used both univariable/multivariable analyses and random forest modeling. The reasons for this dual statistical approach are (1) univariable/multivariable analysis allows the presentation of more traditional statistical measures of association and to allow comparison with published data, (2) random forest is an ensemble method that allows the discovery of variables that are of the most importance as well as allowing the use of decision trees to improve the accuracy of the risk assessment tool, and (3) allow internal cross-validation of the variables in the prediction model.

We conducted univariable analyses using generalized linear modeling. Variables that were significantly associated with 6mNRM (P < .05) on univariable analysis were included in multivariable analysis.

For random forest modeling, selected input variables were known factors associated with transplant survival. Time to death, an outcome variable, and the age-adjusted haemopoietic cell transplant comorbidity index (aaHCT-CI) were included as continuous variables. Gender, conditioning intensity, hematological disease category, donor match, GVHD prophylaxis, delayed neutrophil recovery (>D28), female donor-to-male recipient status, occurrence of aGVHD, donor/recipient CMV serostatus, MAPΔ (≥0.055 vs <0.055), and D7 MAP score (≥0.16 vs <0.16) were included as dichotomous/discrete variables. Data were sampled using the random bootstrapping/bagging method. Based on characteristics of the variables, classification and regression trees were derived. For each decision tree, two-thirds of the data were randomly assigned as the training set, one-third was used for cross-validation. Input variables were ranked using importance measures. The random forest algorithm was used to select variables and develop a risk prediction model. Recursive partitioning was applied to construct survival trees, forming the final risk prediction model, which was subsequently validated on the entire data set.

Statistical software

All analyses were conducted using R version 4.4.3 (R Foundation for Statistical Computing). Inverted Kaplan-Meier survival curves were generated using GraphPad Prism (version 10.2.0). The “survival” and “cmprsk” packages were used for time-to-event and competing risks regression analyses. The “pROC” package was utilized for AUC calculation and threshold optimization, and “survminer” and “ggplot2” were used for graph plotting and data visualization. For the random forest algorithm, the “randomForestSRC” package was applied, whereas “rpart” was used for recursive partitioning in survival tree modeling.

This study was approved by the Alfred Health Ethics Committee (Project ID 78769) conducted in accordance with the Declaration of Helsinki.

Results

Patient characteristics

A total of 101 patients were prospectively recruited between January 2022 and January 2024. Baseline characteristics and outcomes are summarized in Table 1. Median follow-up period was 23.8 months (range, 0.1-36.7). Median age was 60 years (interquartile range, 46-67). The most frequent transplantation indication was acute myeloid leukemia (48/101, 48%), then myelodysplastic syndrome (13/101, 13%), and myeloproliferative neoplasm (13/101, 13%). All stem cells were derived from peripheral blood. Posttransplant cyclophosphamide was the most frequently administered GVHD prophylaxis (83.2%).

Table 1.

Patient baseline characteristics

	Patients (N = 101)
Age, n (%)
<40 years	19 (18.8)
40-49 years	10 (9.9)
50-64 years	44 (43.6)
≥65 years	28 (27.7)
Male sex, n (%)	70 (69.3)
Diagnosis, n (%)
Acute myeloid leukemia	48 (47.5)
Acute lymphoblastic leukemia	12 (11.9)
Myelodysplastic syndrome	13 (12.9)
Myeloproliferative neoplasm	13 (12.9)
Lymphoma	10 (9.9)
Others∗	5 (5)
Refined DRI, n (%)
Low	5 (5)
Intermediate	58 (57.4)
High	36 (35.6)
Very high	0 (0)
aaHCT-CI, n (%)
Low (0)	7 (6.9)
Intermediate (1-2)	57 (56.4)
High (3-4)	26 (25.7)
Very high (≥5)	11 (10.9)
Donor type, n (%)
Matched sibling	28 (27.7)
Matched unrelated donor	43 (42.6)
Haploidentical	30 (29.7)
Conditioning regimen intensity, n (%)
Myeloablative	19 (18.8)
Reduced intensity/nonmyeloablative	82 (81.2)
GVHD prophylaxis, n (%)
Posttransplant cyclophosphamide	84 (83.2)
Methotrexate	9 (8.9)
Alemtuzumab	5 (5)
Others	3 (3)
Antithymocyte globulin use, n (%)	3 (3)
Recipient CMV-positive serostatus, n (%)	72 (71.3)
CMV serostatus (donor/recipient), n (%)
Pos/Pos (D+/R+)	44 (43.6)
Neg/Pos (D−/R+)	28 (27.7)
Pos/Neg (D+/R−)	10 (9.9)
Neg/Neg (D−/R−)	19 (18.8)
Donor/recipient sex, n (%)
Male to male	53 (52.5)
Female to male	17 (16.8)
Male to female	17 (16.8)
Female to female	14 (13.9)
History of previous alloHSCT, n (%)	2 (2)

DRI, disease risk index; n, number; Neg, negative; Pos, positive.

^∗Others: myeloma (n = 3), severe aplastic anemia (n = 2).

Outcomes and nonrelapse causes of death

Cohort NRM was 15.1% (95% confidence interval [CI], 8.87-22.9) at 6 months and 19.7% (95% CI, 12.4-28.3) at 2 years (supplemental Figure 1). Over 2 years, there were 19 NRM events: 16 (84.2%) due to infection, and 1 (5.3%) each from aGVHD, intracranial hemorrhage, and therapy-related acute myeloid leukemia unrelated to the original disease. Among 16 infection-related deaths, 7 (43.8%) were due to bacteremia/sepsis, 2 (12.5%) to invasive fungal disease, 2 (12.5%) to CMV disease, and 5 (31.3%) due to other viral infections.

Only 3 of the 16 infection deaths had preceding grade 2 to 4 aGVHD or were receiving immunosuppression at time of death; thus, 13 of 16 cases (68.4%) fulfilled criteria and were classified as iNRM. Twelve of 13 cases of iNRM occurred within 6 months. The median time to iNRM was 48 days (interquartile range, 23-112). Detailed causes of NRM and infections are provided in supplemental Table 1. No NRM cases were preceded by chronic GVHD.

The incidence of grade 1 to 2 and 3 to 4 aGVHD at 100 days was 10.9% (n = 11) and 14.9% (n = 15), respectively. Disease relapse at 2 years was 31.6% and overall survival at 3 years was 58%.

ST2 and REG3a concentrations in patient serum

Baseline (D0) and D7 serum samples were available for all 101 patients. D14/D21 samples were available for 99 patients; 1 declined collection at both time points, and 1 died at D8 before follow-up. For 5 patients without D21 samples, D14 served as the final follow-up time point (supplemental Figure 2). At D0, mean ST2 was significantly higher in patients receiving myeloablative conditioning than in those receiving reduced-intensity or nonmyeloablative regimens (68 110 vs 53 683 pg/mL; P < .001; Figure 1A). REG3a concentrations did not differ (34.9 vs 19.4 ng/mL; P = .37; Figure 1C). Figure 1B,D shows ST2 and REG3a distributions (mean concentration and quartiles) by 6mNRM status. Mean concentrations across D0, D7, and D14/D21 are detailed in supplemental Table 2. For ST2, concentrations were unchanged in patients with 6mNRM at D14/D21 compared to baseline (59 930 pg/L [D0], 66 789 pg/L [D14/D21]; P = .16; Figure 1B). For REG3a, patients who experienced 6mNRM had significantly increased D14/D21 REG3a concentrations compared to baseline (17.97 ng/L [D0], 51.61 ng/L [D14/D21]; P = .002; Figure 1D). At D14/D21, patients with 6mNRM had significant increases of ST2 and REG3a concentrations compared to those without 6mNRM (ST2: 66 789 pg/L vs 52 292 pg/L, P = .04; REG3a: 51.6 ng/L vs 25.6 ng/L, P = .01; Figure 1B,D).

Figure 1.

ST2 and REG3a concentrations according to alloHSCT timepoints. (A-B) Box and whisker plots showing ST2 and REG3a concentrations at baseline (D0), according to conditioning type. (C-D) Box and whisker plots showing ST2 and REG3a changes across D0, D7 and D21 timepoints; D14 was used if D21 samples were not available.

Patients that experienced an increase in ST2 concentration between D0 and D14/D21 (ST2Δ) had an associated higher risk of tteNRM as assessed with the log-rank test (hazard ratio [HR], 4.58; 95% CI, 1.6-13.2; P = .01). Patients who experienced a REG3a concentration (REG3aΔ) had elevated risk of tteNRM (HR, 13.1; 95% CI, 4.6-37.4; P = .001).

MAP score performance in predicting 6mNRM

MAP score was calculated at D0, D7, and D14/21. Mean MAP scores are presented in supplemental Table 2. Patients who experienced 6mNRM had increased D14/D21 MAP score compared to baseline (0.2119 vs 0.1497; P = .029; Figure 2A). An increase in MAP score of ≥0.055 from D0 to D21 (MAPΔ; number [n] = 99) was strongly associated with tteNRM (HR, 10.43; 95% CI, 3.17-34.3; P < .0001; Figure 2D) using log-rank testing.

Figure 2.

MAP score increase is associated with NRM. (A-B) Box and whisker plots demonstrating MAP score at baseline and follow up timepoints, stratified according to 6mNRM status. (C) Waterfall plot demonstrating the distribution of patients with 6mNRM according to MAPΔ. (D-E) Inverted Kaplan-Meier plots of tteNRM according to MAPΔ and D7 MAP score.

For D7 single time point MAP score value, a cutoff of 0.16, as established in previous publications,⁶ was used. D7 MAP score above/below 0.16 successfully stratified patients (N = 101) into high- and low-risk groups for tteNRM (HR, 6.27; 95% CI, 2.47-15.9; P = .0002; Figure 2E).

The association remains similar when accounting for relapse as a competing risk for both MAPΔ (HR, 11.0; 95% CI, 4.3-28.4; P < .001; supplemental Figure 3A) and D7 MAP score (HR, 6.24; 95% CI, 2.1- 8.1; P < .001; supplemental Figure 3B).

Due to the association of MAPΔ, ST2Δ, REG3aΔ, D7 MAP score, D14/D21 ST2 concentrations and D14/21 REG3a concentrations with increased 6mNRM risk, we compared test predictive performance using ROC curves. Among these, MAPΔ demonstrated the highest discriminatory ability, with an AUC of 0.934. ROC curves and their AUCs are depicted in Figure 3.

Figure 3.

Comparison of test performance using ROC curves. ROC curves for (A) MAPΔ (AUC 0.934), (B) REG3aΔ (AUC 0.867), (C) D7 MAP (AUC 0.779), (D) ST2Δ (AUC 0.748), (E) D14/D21 ST2 (AUC 0.788), (F) D14/D21 REG3a (AUC 0.818).

ROC curve analysis identified the optimal cutoff for the D7 MAP score to predict 6mNRM of 0.16, yielding a sensitivity of 92.9% and a specificity of 67.1%. The optimal cutoff for MAPΔ from D0 to D21 was 0.0559, with a sensitivity of 85.7% and specificity of 89.4%. The inverted Kaplan-Meier curve of tteNRM according to MAPΔ using a threshold of 0.055 is shown in Figure 2D (HR, 10.6; 95% CI, 3.0-37.7; P < .0001).

Among patients admitted to the intensive care unit (n = 24) within 28 days after alloHSCT, MAPΔ ≥0.055 was also associated with 6mNRM (84.6% vs 0% mortality; P < .001).

Univariable and multivariable analysis of risk factors for 6mNRM

To assess strength and magnitude of associations between MAPΔ, the D7 MAP score, and other transplant risk factors with 6mNRM, we performed univariable and multivariable analyses on clinically relevant variables (Table 2). ST2 and REG3a measures were excluded, as MAP includes both biomarkers. On univariable analysis, MAPΔ, the D7 MAP score, aaHCT-CI, delayed neutrophil recovery, and female donor to male recipient were all associated with 6mNRM. However, after adjusting for potential confounders only MAPΔ and female donor-to-male recipient status remained independently associated with 6mNRM. Of note, MAPΔ (≥0.055) and D7 MAP (≥0.16) had variance inflation factor values at 1.56 and 1.30, respectively, suggesting low collinearity. Univariable and multivariable analyses using 6-month iNRM (n = 12) as a dependent variable are shown in supplemental Table 3. Conditioning intensity was not associated with 6mNRM or 6-month iNRM. Insufficient patients had not received in vivo T-cell depletion to allow meaningful statistical comparisons.

Table 2.

Univariable and multivariable logistic regression analyses of factors associated with 6mNRM (all-cause 6mNRM)

	Univariate analysis		Multivariate analysis
	OR (95% CI)	Univariate P value	OR (95% CI)	Multivariate P value
aaHCT-CI	2.35 (1.18-4.85)	.016∗	1.51 (0.4-6.1)	.546
MAPΔ ≥0.055	45 (10.5-318.7)	<.001∗	24.4 (2.59-624)	.013∗
D7 MAP >.16	27.5 (5.15-511)	.002∗	7.4 (0.64-276)	.162
CMV status donor/recipient
Pos/Neg	2.12 (0.22-20.5)	.489
Neg/Pos	3.4 (0.73-25)	.153
Pos/Pos	0.62 (0.09-5.03)	.620
Neutrophil recovery >28 days	33.2 (2.75-793)	.007∗	6.5 (0.047-150)	.515
Prophylaxis
PTCy use	1.37 (0.33-9.36)	.7
Grade 2 to 4 aGVHD	0.32 (0.04-1.25)	.150
Male sex	0.87 (0.28-3)	.810
Conditioning intensity
Myeloablative	1	NA
Reduced intensity/nonmyeloablative	3.7 (0.67-69)	.220
Donor type
Matched sibling	1	NA
Matched unrelated donor	2.97 (0.68-20)	.190
Haploidentical	2.6 (0.51-19)	.279
Female donor to male recipient	3.74 (1.1-12)	.030∗	13.1 (1.47-303)	.038∗

D7 MAP, MAP score at D7 after alloHSCT; NA, not applicable.

^∗Denotes statistical significance.

Random forest modeling to determine variables of greatest importance

To develop a predictive model for 6mNRM, we performed random forest modeling using key transplant risk factors. The top 5 predictors were MAPΔ (feature importance: 0.1870), D7 MAP >.16 (0.0449), CMV donor/recipient serostatus (0.005), aaHCT-CI (0.003), and female donor to male recipient (0.001). This returns an acceptable out-of-bag error of 13.4%, which estimates the error rate (1-accuracy) that the training approach has for the test data from the same patient set. Feature importance comparison is depicted in supplemental Figure 4.

To develop a risk prediction model for 6mNRM, we used the top 5 variables to generate a decision tree using recursive partitioning shown in supplemental Figure 5. Node or bin sizes had at least 5 individuals to sufficiently visualize decision tree levels and identify the contribution of each variable to each branch. The performance metrics of the decision tree are shown in supplemental Figure 5. To simplify the model and reduce the possibility of introducing chance with small numbers in the end points, only the first 2 levels of the decision tree were selected. The predictive model is illustrated in Figure 4A, whereas the separation of NRM occurrence based on risk groups according to this model is demonstrated in Figure 4B (n = 99). The model returned a root node error (percent of correctly sorted records) of 80.2% and cross-validation error rate (correctly sorted records on internal cross-validation) of 84.5%, which suggested satisfactory predictive performance. We compared it to the aaHCT-CI, widely used to inform NRM risk, and stratified patients into 3 risk groups (Figure 4C). The high-risk group of our model had 73.3% (95% CI, 40.7-89.9) NRM at 6 months, compared to the aaHCT-CI model that had 28.3% (95% CI, 14.4-43.8) NRM at 6 months. Additional supplemental analysis using Gray’s test to account for relapse as competing risk and curves are shown in supplemental Figure 6A-B. Using tteNRM as an outcome, the high-risk group compared to the low-risk group in our model had a significantly increased risk of NRM with a HR of 62.7 (95% CI, 8.3-471; P < .001) in the competing risk model, compared to the aaHCT ≥3 group that had a HR of 2.8 (95% CI, 1.1-6.8; P = .024) (supplemental Figure 6A-B).

Figure 4.

Risk assessment tool design and performance. (A) Flow chart of the proposed risk assessment tool that stratifies patients into low, intermediate and high-risk groups. (4B-C) Inverted Kaplan-Meier curves of NRM according to (B) the proposed risk assessment tool involving MAPΔ and (C) aaHCT-CI.

Discussion

This study shows that MAPΔ changes within the first 3 weeks after alloHSCT are strongly associated with 6mNRM. Infection was the most common cause of NRM, occurring independently of acute GVHD, highlighting the relevance of MAPΔ to non-GVHD complications. MAPΔ also stratified 6mNRM risk among patients admitted to intensive care unit within the first month of alloHSCT. In logistic regression analyses of 6mNRM, we compared MAPΔ with a single D7 MAP score for predicting 6mNRM. Both stratified patients into high- and low-risk groups, but only MAPΔ independently predicted mortality in multivariable analyses. MAPΔ also ranked as the most important variable in random forest analyses. Additionally, we observed higher ST2 levels in patients who received myeloablative conditioning compared to reduced-intensity or nonmyeloablative regimens. This suggests that baseline ST2, and by extension, MAP scores, are influenced by conditioning intensity and other treatment-related factors. Thus, a dynamic, longitudinal measure like MAPΔ provides a more accurate risk assessment than a single time point.

In our cohort, clinically significant and severe acute GVHD incidence was comparable to that reported by other transplant centers (overall, 25.7%; grade 3-4, 14.9%; gastrointestinal/liver involvement, 12.8%). However, associated NRM was low, with only 3 cases recorded. This favorable outcome is likely attributable to the high proportion of patients who received posttransplant cyclophosphamide (PTCy). Existing literature supports this observation, indicating that PTCy-based GVHD prophylaxis is associated with higher rates of treatment-responsive GVHD and reduced NRM.^12,13 Instead, we observed a high incidence of iNRM, including deaths from bacterial, fungal, and viral pathogens, consistent with prior retrospective studies.1, 2, 3 Notably, MAP score reflects gastrointestinal damage, consistent with most cases of bacterial sepsis being caused by gastrointestinal-associated microbes, related to bacterial translocation.¹⁴ An Australian registry study including various GVHD prophylaxis regimens reported infection-related mortality (IRM) incidences of 6.2% at D100 and 9.8% at 1-year after alloHSCT among adult patients.¹⁵ Irene et al¹⁶ conducted a retrospective single-centre study of 117 adult patients receiving PTCy-based prophylaxis for haploidentical, matched sibling, and unrelated donors and reported a 6-month IRM rate of 12%, with bacterial infections accounting for 42% of cases. This was comparable to our observed 6-month iNRM incidence (11.8%), which was likewise predominated by bacterial infections. However, the Australian registry and Irene et al classified patients with GVHD within the IRM group, a definition that differs from the iNRM definition that we used in our study.^11,15,16 As PTCy gains popularity for GVHD prophylaxis in haploidentical and nonhaploidentical settings, evidence suggests identifying patient subgroups most vulnerable to IRM indicates who benefits from intensified infection prevention and prompt, aggressive management.

We demonstrate that MAPΔ strongly predicts mortality independent of GVHD, with substantial infectious deaths despite standard antimicrobial prophylaxis. To enhance risk stratification, we developed a predictive model using random forest analysis. This tool identifies an exceptionally high-risk subgroup for 6mNRM by integrating MAPΔ, D7 MAP, and aaHCT-CI score. We demonstrate that MAPΔ and D7 MAP have low or at most moderate collinearity from their variance inflation factor scores (1.56 and 1.30, respectively). Patients with both a MAPΔ ≥0.055 and an aaHCT-CI ≥2 experienced a >70% risk of 6mNRM. Pending external validation, we propose that patients with an aaHCT-CI ≥2 constitute a target population for serial MAP measurement.

NRM risk stratification following alloHSCT traditionally relies on clinical tools including the HCT-CI, its age-adjusted version, the European Society for Blood and Marrow Transplantation risk score, and the simplified comorbidity index, which integrate baseline comorbidities and donor-related factors to estimate mortality risk.17, 18, 19 More recently, integrating biomarkers is a promising approach to enhance model predictive accuracy. Tools combining clinical parameters and biomarker data may offer superior prognostication compared to either approach alone. The MAP score, based on serum biomarkers ST2 and REG3a, has primarily been used to stratify NRM risk at aGVHD onset and has shown improved predictive performance when combined with clinical symptoms.²⁰ Although elevated early posttransplant levels of ST2, REG3a, and MAP score have been associated with increased NRM risk, they have not previously been incorporated into a comprehensive early posttransplant risk model. To our knowledge, our model is the first to integrate the MAP score with aaHCT-CI to predict early NRM not limited to aGVHD.

There are several strategies that have been explored to reduce infectious mortality risk in this vulnerable patient group. Widespread antifungal prophylaxis in alloHSCT, with anti-CMV prophylaxis or preemptive therapy, has significantly reduced life-threatening infections in transplant recipients. Only 2 patients in our NRM cohort died from fungal infections, likely reflecting the routine use of antimold prophylaxis, therapeutic drug monitoring for all patients receiving antifungals and aggressive investigation of patients with suspected invasive fungal infection at our institution. Consequently, the relative contribution of pathogens to iNRM is expected to vary according to institutional prophylaxis strategies. IV immunoglobulin (IVIG) and antimicrobial prophylaxis have been studied in patients with hematological malignancies, showing benefit in reducing infections but not mortality. A systematic review by Chai et al found IVIG use was associated with significantly fewer documented infections (risk ratio [RR], 0.72; 95% CI, 0.54-0.96), including serious infections in 1 study (RR, 0.34; 95% CI, 0.14-0.81), without affecting all-cause mortality.²¹ In the alloHSCT setting, randomized controlled trials involving IVIG have not shown differences in infection rates or survival.^22,23 The use of antibiotics, particularly fluoroquinolones (FQs), remains controversial due to risk of promoting antibiotic resistance. A 2018 meta-analysis including 2 randomized controlled trials and 12 observational studies found that FQ prophylaxis was associated with a fewer bloodstream infections (pooled OR, 0.57; 95% CI, 0.43-0.74) and fewer episodes of neutropenic fever (pooled OR, 0.32; 95% CI, 0.20-0.50), with no impact on mortality (pooled OR, 1.01; 95% CI, 0.73-1.41).²⁴ Both IVIG and antibiotic prophylaxis reduce significant bacterial infections without affecting mortality. This may relate to the unselected nature of the populations studied, where overall effects are diluted by nonbenefiting low-risk patients. However, in high-risk populations, strategies including targeted antimicrobials and IVIG may improve outcomes. Our risk assessment tool could inform the design of interventional trials targeting the highest-risk individuals, most of them benefitting from enhanced infection prevention. In addition, this tool could guide targeted antibiotic prophylaxis and IVIG in high-risk groups instead of universal FQ prophylaxis. This would minimize antibiotic resistance and other adversity.

Our study is limited by a relatively small sample size, precluding validation in a separate group. However, the random forest ensemble method mitigates this, resulting in an acceptable 13.4% error rate. Nevertheless, future larger studies are warranted for further validation. There is the possibility of residual confounding by covariates that may influence iNRM, for example mucositis severity and Karnofsky Performance Scale. However, the covariates chosen for this analysis were known factors associated with alloHSCT outcomes, hence we believe sufficient covariates were included for model accuracy.

Conclusion

In summary, MAPΔ was strongly associated with 6mNRM, primarily due to infections. We demonstrate that changes in the MAP score (MAPΔ) outperform single time point MAP score measurements in predicting 6mNRM, independent of GVHD. Based on this, we developed a risk stratification tool that effectively identifies patients at significantly elevated risk for IRM. This tool may be used to identify the most vulnerable patients who could be targeted with more intensive clinical monitoring, additional antimicrobial prophylaxis, or reductions in immunosuppression/regimen intensity, where possible. However, further prospective validation in larger, independent cohorts are needed to confirm its clinical utility and ultimately interventional trials are required.

Conflict-of-interest disclosure: J.R. holds stock in Novartis AG, Sandoz AG, and Alcon as a former employee of Novartis; and his current employer, Alfred Hospital, receives payments to recoup his inputs to trials funded by Janssen, Novartis, AbbVie, GSK, Antengene, and Bristol Myers Squibb. The remaining authors declare no competing financial interests.