Immunologic clustering of donor-specific antibodies and clinical outcomes in kidney transplant recipients

Abstract

BACKGROUND

Donor-specific antibodies (DSA) against human leukocyte antigens (HLA) are associated with increased immunologic risk in kidney transplant recipients (KTR). However, outcomes among DSA-positive patients are highly variable. Traditional markers, such as DSA class and mean fluorescence intensity (MFI), often fail to capture the multidimensional nature of immunologic risk.

AIM

This retrospective study aims to stratify DSA-positive KTR into immunologic risk groups using unsupervised machine learning and to assess the relationship between antibody profiles and early post-transplant outcomes.

METHODS

During the study period, 260 kidney transplants were performed at King Fahad Armed Forces Hospital in Saudi Arabia, of which 151 recipients were DSA-positive. Clustering analysis was performed in 112 DSA-positive recipients with complete immunologic and follow-up data. Key variables used for clustering included DSA class, number of DSA, cumulative MFI, early graft function (estimated glomerular filtration rate at 4 months), and HLA mismatch. K-means clustering was applied to identify distinct immunologic risk groups. Clinical outcomes, including acute rejection, graft loss, delayed graft function, and mortality, were compared across clusters.

RESULTS

Three distinct clusters were identified: High-risk (n = 58) with both Class I and II DSA, high DSA numbers and MFI, and the highest rates of acute rejection (6.9%), graft loss (6.9%), and mortality (5.2%); intermediate-risk (n = 34) with predominantly Class II DSA, moderate MFI, and mild graft loss and mortality; and low-risk (n = 20) with isolated Class I DSA, low DSA numbers and MFI, and excellent early graft function with minimal complications. Clusters demonstrated a clear gradation of immunologic burden, which corresponded with early post-transplant outcomes. Principal component analysis showed clear separation among the three groups.

CONCLUSION

Unsupervised clustering effectively stratifies DSA-positive KTR into clinically meaningful risk groups. High-risk recipients with broad and strong DSA profiles are more likely to experience early adverse outcomes. Low-risk recipients have favorable early graft function. This data-driven approach may help guide individualized monitoring, donor selection, and immunosuppression strategies for DSA-positive KTR.

Key Words: Immunological clustering; Donor-specific antibodies; Renal allograft outcomes; Machine learning; Human leukocytes antigen

Core Tip: Donor-specific antibodies (DSA) in kidney transplant recipients (KTR) are associated with variable immunologic risk, which is not always well defined by DSA class or mean fluorescence intensity alone. In this study, an unsupervised machine learning approach was used to group DSA-positive recipients based on antibody characteristics and transplant-related factors. These groups showed apparent differences in early post-transplant risk profiles. Patients with broader and stronger antibody responses were more likely to experience adverse early outcomes, whereas those with a limited antibody burden had favorable graft function. This approach may help refine risk assessment and support more individualized management of DSA-positive KTR.

INTRODUCTION

Donor-specific antibodies (DSA) are directed against human leukocyte antigen (HLA) and may cause acute antibody-mediated rejection, chronic antibody-mediated rejection, and graft loss[1]. Traditionally, the presence of DSA in kidney transplant recipients (KTR) is considered a higher immunological risk[2]. However, outcomes within this group are highly variable[3]. Immunological risk assessment among DSA-positive patients is based on parameters such as the mean fluorescence intensity (MFI) of antibodies and DSA class (Class I or II)[4], as well as complement-binding ability[5]. These parameters affect the renal allograft outcomes. While these markers provide valuable information, they are often evaluated in isolation and may not capture the combined influence of antibody strength, breadth, and antigen specificity. Moreover, MFI values are semi-quantitative and subject to inter-laboratory variability, which limits their predictive accuracy[6].

The immunologic profile of DSA-positive recipients is complex and involves multiple factors. These include the number of antibodies, the specific HLA targets they recognize, and the potential for cross-reactivity with other antigens. Such variations help explain why two patients with similar antibody strength or class can have very different graft outcomes[7]. Recently, researchers have begun using data-driven approaches, including clustering and machine learning, to better understand this complexity[8-10]. Cluster analysis is a statistical method that groups patients or samples based on shared characteristics, creating subgroups that are more similar within the group than between groups. In kidney transplantation, this approach helps identify patient subpopulations with distinct immunologic or clinical profiles, even when conventional measures fail to explain differences in outcomes. By revealing these patterns, cluster analysis provides a practical framework to better understand the complexity of transplant immunology and its impact on graft outcomes. Studies using machine-learning-based clustering have provided new insights into the heterogeneity of kidney transplantation. For example, unsupervised analyses of large retransplant cohorts have revealed distinct patient subgroups that differ in sensitization levels, donor types, and clinical outcomes. In some studies, moderately sensitized patients were found to have the poorest graft and patient survival[8]. Other analyses of kidney biopsies have refined the Banff classification, identifying six rejection phenotypes and clarifying previously ambiguous categories, which correlated more strongly with graft failure[9]. Similarly, in studies of highly sensitized recipients (panel reactive antibodies ≥ 98%), two contrasting clusters were observed: One group of younger, male repeat recipients with higher rejection rates and poorer graft survival, and another group of older, first-time female recipients who generally had better outcomes[10]. Taken together, these findings show that examining the immune system from multiple angles, including antibody strength, breadth, and specificity, as well as patient and donor factors, helps clarify previously “unexplained” differences in transplant outcomes and supports more personalized risk assessment.

Studies based on clustering analyses are lacking in Saudi Arabia. To date, no study has been conducted to cluster DSA and demonstrate its immunological impact on graft outcomes. Conducting such analyses will provide valuable insights into region-specific risk profiles and outcomes. To address this gap, we aimed to stratify KTR with DSA⁺ into distinct immunologic risk groups using unsupervised clustering. This approach was designed to identify meaningful patient subgroups based on the strength and breadth of antibody sensitization, graft function, and immunologic mismatch, without predefining clinical outcomes.

MATERIALS AND METHODS

Study design and population

This retrospective observational study was conducted at the Renal Diseases and Transplantation Center, King Fahad Armed Forces Hospital, Jeddah, Saudi Arabia. The institutional ethics review committee approved the study in accordance with the principles outlined in the Declaration of Helsinki. Informed consent was obtained from the study participants or their next of kin.

During the study period (2021-2024), 260 adult kidney transplants were performed at our center. Among them, 151 recipients (58.1%) had confirmed donor-specific HLA antibodies (DSA), identified using single-antigen bead (SAB) (Luminex) assays.

All adult KTR (≥ 18 years) with confirmed donor-specific HLA antibodies (DSA⁺), identified using SAB (Luminex) assays, were eligible for inclusion. All included recipients had negative pre-transplant crossmatches, as per standard immunologic evaluation. Of the 112 DSA-positive KTR included in the clustering analysis, 104 (92.9%) received kidneys from living donors. All living donors were either biologically related to the recipients (up to the 4^th degree) or were the recipients’ spouses. No unrelated donors outside these categories were included, and all donors were ≥ 18 years of age. Patients were required to have complete data for key clustering variables, including antibody MFI values (Class I and II), number of DSAs, estimated glomerular filtration rate (eGFR) at 4 months, HLA mismatch, and DSA class type. Recipients were excluded if they underwent multi-organ transplantation, lacked confirmed DSA characterization, had missing data for any clustering feature, experienced primary graft non-function due to surgical causes, or were lost to follow-up before 4 months post-transplant. Of the 151 DSA-positive recipients, 112 met the inclusion criteria and were included in the clustering analysis. Thirty-nine DSA-positive recipients were excluded due to incomplete immunologic characterization or missing follow-up data.

Study variables

The study variables included various demographic, clinical, immunological, and outcome parameters. Demographic variables included recipient and donor age and gender, as well as donor type (living or deceased). Clinical variables included pre-transplant diabetes mellitus, history of previous kidney transplantation, length of hospital stay, and the maintenance immunosuppressive regimen. Immunological variables comprised the presence of DSA, type of induction therapy (antithymocyte globulin or basiliximab), sensitizing events such as pregnancy, blood transfusion, or plasma exchange, and the use of intravenous immunoglobulin (IVIG). Outcome variables included acute rejection, delayed graft function (DGF), graft loss, mortality, and infections with cytomegalovirus (CMV) or BK virus. Continuous variables analyzed included recipient and donor age, length of hospital stay, serum creatinine at four months, and eGFR at four months.

HLA-typing and antibody detection

All kidney transplant candidates were screened for HLA antibodies using SAB assays. ABO compatibility and HLA matching were confirmed for every donor–recipient pair. HLA typing for both donors and recipients was performed using the reverse sequence-specific oligonucleotide technique on the Luminex 3D platform (Luminex Corporation, Austin, TX, United States). Genomic DNA was extracted from peripheral blood samples, and the loci analyzed included HLA-A, -B, -C, -DRB1, -DPB1, and -DQB1. Typing was conducted at low to intermediate resolution (first-field level). Antigen-level equivalence was used to assign match grades in accordance with standard clinical laboratory practice. In selected cases with ambiguous alleles, two-field confirmation was performed. Hybridization and detection procedures were performed according to the manufacturer’s instructions, and fluorescence data were analyzed using HLA Fusion software (One Lambda, Canoga Park, CA, United States). Anti-HLA antibodies were identified using the SAB assay on the Luminex 3D platform. Positive and negative controls recommended by the manufacturer were included in all assays. Samples showing non-specific or denatured antigen binding were retested to ensure accurate identification of antibodies.

Induction and maintenance immunosuppression

Our center primarily uses induction with anti thymocyte globulin (ATG). Basiliximab is occasionally used in patients older than 60, those with leukopenia, those who had complete HLA matching, or recipients with no immunoglobulin G antibodies against Epstein-Barr virus. All KTR were treated with rabbit ATG (Thymoglobulin). It was administered at a dose of 1.5 mg/kg for four doses, resulting in a total dose of 4-6 mg/kg, depending on white blood cell and platelet counts. We used a triple regimen including tacrolimus (FK), mycophenolate mofetil (MMF), and prednisolone (PRED) as maintenance therapy. Cyclosporine was used occasionally as needed based on clinical considerations. Methylprednisolone 500 mg was administered in the operating theatre on day zero. This is followed by 250 mg on day 1 and 125 mg on day 2. After the third dose of methylprednisolone, PRED was given 50 mg on day 3. It is then reduced by 10 mg per day till the patient reaches 20 mg. All patients were discharged on 20 mg PRED (Pred), which was then lowered to 5 mg per day over 1 month. MMF was administered at a dose of 500 mg during ATG. After ATG completion, the dose was increased to 1000 mg twice daily. All patients received valganciclovir 450 mg for 3-6 months, depending on their CMV serological status. Patients received prophylaxis against Pneumocystis Jirovecii with cotrimoxazole (trimethoprim-sulfamethoxazole, 80/400 mg once daily) for 12 months. They also received oral nystatin suspension (100000 units/mL; 1 mL, swished and swallowed four times daily) for 1 month. Rejection episodes were classified according to the Banff criteria in use at the time of biopsy (Banff 2019 or Banff 2022)[11,12].

Unsupervised clustering of donor-specific antibody profiles

Patients with confirmed DSA datasets were analyzed. Variables were selected based on their clinical relevance to immunologic risk and post-transplant graft outcomes. Variables of interest included the MFI of Class I and Class II antibodies, number of distinct DSAs, eGFR as a marker of early graft function, degree of HLA mismatch, and DSA class type (Class I, Class II, or both). Before clustering, data preprocessing and encoding were done to ensure comparability across variables. Cumulative DSA strength was calculated as the sum of each patient's individual DSA MFI values, reflecting the total antibody burden. Class-specific MFIs (Class I and Class II) were analyzed separately to characterize immunologic profiles. All MFI variables were log-transformed and standardized prior to clustering to reduce skewness and ensure comparability across features. The DSA type was ordinally encoded to reflect the immunologic hierarchy (Class I < Class II < Both). All continuous features were standardized using Z-score normalization to equalize their contribution during clustering. Patients with missing values were excluded using listwise deletion to maintain internal consistency. Clustering was then performed using the K-means algorithm. This partition-based unsupervised machine learning method assigns each observation to one of k clusters based on feature similarity by minimizing intra-cluster variance. The optimal number of clusters (k) was determined using the elbow and silhouette methods. This is to achieve the best balance between compactness and separation. The resulting clusters were subsequently used to compare immunologic features and clinical outcomes among KTR.

Statistical analysis

All analyses were performed using Python in a Jupyter Notebook environment. Categorical variables are presented as counts and percentages. Continuous variables are reported as medians with interquartile ranges (IQR). Comparisons of continuous variables across clusters were conducted using the Kruskal-Wallis test. In contrast, categorical variables were compared using the χ² test or Fisher’s exact test when expected cell counts were small. Unsupervised clustering was performed using the K-means algorithm on standardized and preprocessed features, including DSA class and strength (MFI), number of DSAs, HLA mismatches, and early graft function (eGFR). The optimal number of clusters was determined using the Elbow method and silhouette scores. Cluster separation was visualized with principal component analysis (PCA) and heatmaps. Patterns of antibody combinations were further explored using UpSet plots. Clinical outcomes, including acute rejection, graft loss, DGF, and mortality, were summarized by cluster. Time-to-event analyses for a composite endpoint were performed using Kaplan-Meier survival curves with the log-rank test. A P value < 0.05 was considered statistically significant.

We attempted to minimize potential sources of bias in several ways. All consecutive DSA-positive KTR during the study period were screened using predefined inclusion and exclusion criteria to reduce selection bias. Only patients with complete immunologic characterization and follow-up data were included in the clustering analysis to ensure dataset consistency. Laboratory testing for HLA typing and DSA detection was performed using standardized Luminex SAB assays, and rejection episodes were classified according to the Banff criteria in use at the time of biopsy to ensure uniform outcome definitions. Clustering was conducted without including outcome variables to avoid outcome-driven grouping. Continuous variables were log-transformed and standardized prior to analysis to ensure balanced contribution of each feature.

RESULTS

Patient demographics and baseline characteristics

Of the 151 DSA-positive recipients identified during the study period, 112 fulfilled the predefined inclusion criteria and were included in the clustering analysis. Thirty-nine patients were excluded because of incomplete immunologic characterization and/or missing follow-up data and were therefore not eligible for risk-stratification analysis. Patients were followed for a minimum of 4 months post-transplant, with available follow-up extending up to 24 months. Graft function was assessed at 4 months, 1 year, and 2 years. Time-to-event analyses were performed during the available follow-up period. Among 112 DSA-positive KTR, 51.8% were male and 48.2% female. Donors were mainly male (68.8%) and living (92.9%). The majority received ATG induction (96.4%) and maintenance therapy with FK, MMF, and PRED. Most patients had no sensitizing events (92.0%). Only 2.7% had a previous transplant. IVIG was administered in 59.8% of cases. The median recipient age was 46.0 years, and the median donor age was 35.0 years. Median hospital stay was 9.0 days. Kidney function remained stable, with median creatinine levels of 90.0, 83.0, and 82.0 μmol/L, and median glomerular filtration rate (GFR) levels of 78.0, 86.0, and 86.5 mL/minute at 4 months, 1 year, and 2 years, respectively. Table 1 shows patients' demographic and baseline characteristics.

Table 1 Patient demographics and baseline characteristics.

Variable	Category	n (%)
Recipient gender	Female	54 (48.2)
	Male	58 (51.8)
Donor gender	Female	35 (31.2)
	Male	77 (68.8)
Donor type	Deceased	8 (7.1)
	Living	104 (92.9)
DSA presence	Present	112 (100.0)
Induction therapy	ATG only	108 (96.4)
	Basiliximab only	4 (3.6)
Sensitizing events	None	103 (92.0)
	Pregnancy	5 (4.5)
	Blood Transfusion	4 (3.5)
Maintenance therapy	CSA + MMF + PRED	1 (0.9)
	FK + MMF + PRED	111 (99.1)
Pre-transplant desensitization	No	60 (53.6)
	Yes	52 (46.4)
Previous transplant	No	109 (97.3)
	Yes	3 (2.7)
IVIG use	No	45 (40.2)
	Yes	67 (59.8)
Variable	Median	IQR/unit
Recipient age (years)	46.0	35.0-60.0
Donor age (years)	35.0	29.0-43.2
ATG dose (mg/kg)	5.0	4.0-5.0
Length of stay (days)	9.0	8.0-12.0
Creatinine-4 months (μmol/L)	90.0	72.0-112.0
Creatinine-1 year (μmol/L)	83.0	72.5-107.0
Creatinine-2 years (μmol/L)	82.0	69.2-104.5
eGFR-4 months (mL/minute)	78.0	62.5-96.2
eGFR-1 year (mL/minute)	86.0	67.0-99.5
eGFR-2 years (mL/minute)	86.5	72.5-102.0

DSA: Donor-specific antibody; ATG: Anti-thymocyte globulin; CSA: Cyclosporine A; FK: Tacrolimus; MMF: Mycophenolate mofetil; PRED: Prednisolone; IVIG: Intravenous immunoglobulin; eGFR: Estimated glomerular filtration rate; IQR: Interquartile range.

Optimal cluster number and risk stratification

We used unsupervised K-means clustering. Among 112 DSA-positive patients, we identified three distinct immunologic risk clusters. By using the elbow method and silhouette scores, we determined that k = 3 is the optimal number of clusters (Figure 1), resulting in the most distinct and internally consistent grouping.

Figure 1 Evaluation of cluster number using the elbow and silhouette methods. A: The elbow method, highlighting a sharp decline in within-cluster variance at k = 3; B: Silhouette scores across k values, peaking at k = 3. Together, these findings support the selection of three clusters for stratifying immunologic risk.

Cluster profiles and clinical features

We identified three distinct clusters. The high-risk cluster (n = 58) was characterized by combined Class I and Class II DSA, with a median of 5 antibodies (IQR 2.0-7.8) and the highest cumulative MFI (median 3811; IQR 2366.5-6998.8). Median Class I and Class II MFIs were 1174 and 1966, respectively. This group demonstrated broad antibody reactivity and the highest observed rates of acute rejection, graft loss, and mortality. Cluster 1 (intermediate risk, n = 34) exhibited predominantly Class II sensitization with moderate MFI, mild graft loss and mortality, and no early acute rejection. Cluster 2 (low risk, n = 20) consisted of patients with isolated Class I sensitization, low DSA numbers and MFI, excellent graft function, and no significant clinical complications. Collectively, these clusters demonstrate meaningful differences in immunologic burden and post-transplant outcomes, supporting stratification into high-, intermediate-, and low-risk groups (Table 2).

Table 2 Clinical interpretation and labeling of the three clusters based on immunologic and outcome profiles.

Cluster	Immunologic profile	Risk interpretation	Key characteristics
0	Combined Class I + II sensitization	High risk	Highest DSA number, high MFI, broad reactivity, highest rejection/mortality
1	Predominantly Class II sensitization	Intermediate risk	Moderate MFI, mild graft loss/mortality, no acute rejection
2	Class I only sensitization	Low risk	Low MFI, low DSA number, no graft loss or mortality

K-means clustering completed with 3 clusters, in a total of 112. Cluster 0 had 58 patients. Cluster 1 had 34 patients. Cluster 2 had 20 patients. DSA: Donor-specific antibody; MFI: Mean fluorescence intensity.

Patient demographics and baseline characteristics by cluster

High-risk patients (cluster 0, n = 58) were characterized by a combination of Class I and Class II DSA. They had the highest DSA numbers and MFI with broad antibody reactivity. Most recipients received kidneys from living donors (89.7%), and induction therapy consisted exclusively of ATG (100%). This group had the highest rates of acute rejection (6.9%), graft loss (6.9%), and mortality (5.2%). Median GFR at 4 months was 76.5 mL/minute.

Intermediate-risk patients (cluster 1, n = 34) predominantly had Class II DSA with a median MFI of 2277 and a median DSA count of 3. Most donors were living (33/34, 97.1%), and induction was mainly performed with ATG (31/34, 91.2%). No acute rejection occurred; however, graft loss (2/34, 5.9%) and mortality (2/34, 5.9%) were observed, with a median GFR of 78 mL/minute.

Low-risk patients (cluster 2, n = 20) had isolated Class I DSA, low DSA numbers (median, 1), and a Class I MFI (MFI, 686), as well as excellent early graft function (median GFR, 92 mL/minute). Only 1/20 (5.0%) had acute rejection, and there was no graft loss or mortality.

Overall, the three clusters showed clear differences in immunologic burden, DSA characteristics, and early post-transplant outcomes. This supports their stratification into high-, intermediate-, and low-risk groups. Table 3 shows patients' demographics and baseline characteristics by cluster.

Table 3 Patient demographics and baseline characteristics by cluster.

Risk group	Variable	Category	n (%)
Categorical variables
High risk	Recipient gender	Male	30 (51.7)
		Female	28 (48.3)
	Donor gender	Male	37 (63.8)
		Female	21 (36.2)
	Donor type	Living	52 (89.7)
		Deceased	6 (10.3)
Intermediate risk	Recipient gender	Male	17 (50.0)
		Female	17 (50.0)
	Donor gender	Male	27 (79.4)
		Female	7 (20.6)
	Donor type	Living	33 (97.1)
		Deceased	1 (2.9)
Low risk	Recipient gender	Male	11 (55.0)
		Female	9 (45.0)
	Donor gender	Male	13 (65.0)
		Female	7 (35.0)
	Donor type	Living	19 (95.0)
		Deceased	1 (5.0)
High risk	DSA presence	Present	58 (100.0)
	Induction therapy	ATG only	58 (100.0)
	Sensitizing events	None	53 (91.4)
		Pregnancy	3 (5.2)
		Blood transfusion	2 (3.4)
Intermediate risk	DSA presence	Present	34 (100.0)
	Induction therapy	ATG only	31 (91.2)
		Basiliximab only	3 (8.8)
	Sensitizing events	None	31 (91.2)
		Pregnancy	3 (8.8)
Low risk	DSA presence	Present	20 (100.0)
	Induction therapy	ATG only	19 (95.0)
		Basiliximab only	1 (5.0)
	Sensitizing events	None	19 (95.0)
		Blood Transfusion	1 (5.0)
High risk	Maintenance therapy	FK + MMF + PRED	57 (98.3)
		CSA + MMF + PRED	1 (1.7)
	Diabetes pre-transplant	No	31 (53.4)
		Yes	27 (46.6)
	Previous transplant	No	56 (96.6)
		Yes	2 (3.2)
	IVIG use	No	18 (31.0)
		Yes	40 (69.0)
Intermediate risk	Maintenance therapy	FK + MMF + PRED	34 (100.0)
	Diabetes pre-transplant	No	16 (47.1)
		Yes	18 (52.9)
	Previous transplant	No	34 (100.0)
	IVIG use	No	15 (44.1)
		Yes	19 (55.9)
Low risk	Maintenance therapy	FK + MMF + PRED	20 (100.0)
	Diabetes pre-transplant	No	13 (65.0)
		Yes	7 (35.0)
	Previous transplant	No	19 (95.0)
		Yes	1 (5.0)
	IVIG use	No	12 (60.0)
		Yes	8 (40.0)
High risk	Acute rejection	No	54 (93.1)
		Yes	4 (6.9)
	Graft loss	No	54 (93.1)
		Yes	4 (6.9)
	Delayed graft function	No	56 (96.6)
		Yes	2 (3.4)
	Mortality	No	55 (94.8)
		Yes	3 (5.2)
	CMV infection	No	56 (96.6)
		Yes	2 (3.4)
	BKV infection	No	54 (98.2)
		Yes	1 (1.8)
Intermediate risk	Acute rejection	No	34 (100.0)
	Graft loss	No	32 (94.1)
		Yes	2 (5.9)
	Delayed graft function	No	32 (94.1)
		Yes	2 (5.9)
	Mortality	No	32 (94.1)
		Yes	2 (5.9)
	CMV infection	No	33 (97.1)
		Yes	1 (2.9)
	BKV infection	No	29 (96.7)
		Yes	1 (3.3)
Low risk	Acute rejection	No	19 (95.0)
		Yes	1 (5.0)
	Graft loss	No	20 (100.0)
	Delayed graft function	No	20 (100.0)
	Mortality	No	20 (100.0)
	CMV infection	No	20 (100.0)
	BKV infection	No	19 (100.0)
Continuous variables
Risk group	Variable	Median	IQR/unit
High risk (years)	Recipient age	46.5	31.5-59.8
	Donor age	35.5	29.0-44.0
Intermediate risk (years)	Recipient age	51.5	39.5-62.8
	Donor age	35	29.0-43.2
Low risk (years)	Recipient age	42.5	36.5-52.8
	Donor age	32.5	27.2-39.2
High risk
	Length of stay (days)	9	8.0-12.0
	Creatinine 4 months (μmol/L)	94	76.0-114.0
	eGFR 4 months (mL/minute)	76.5	63.2-88.0
Intermediate risk
	Length of stay (days)	10	8.0-11.0
	Creatinine 4 months (μmol/L)	88	70.5-110.8
	eGFR 4 months (mL/minute)	78	59.5-97.5
Low risk	Length of stay (days)	9	7.0-12.5
	Creatinine 4 months (μmol/L)	77.5	69.8-103.8
	eGFR 4 months (mL/minute)	92	66.2-109.5

DSA: Donor-specific antibody; ATG: Anti-thymocyte globulin; FK: Tacrolimus; MMF: Mycophenolate mofetil; CSA: Cyclosporine A; PRED: Prednisolone; IVIG: Intravenous immunoglobulin; eGFR: Estimated glomerular filtration rate; IQR: Interquartile range; LOS: Length of stay; AR: Acute rejection; DGF: Delayed graft function; CMV: Cytomegalovirus; BKV: BK virus.

Comparison of cluster features

A univariate analysis was conducted to compare clinical and immunological characteristics across the three clusters. Recipient gender was evenly distributed, with males representing 51.7% of the high-risk group, 50.0% of the intermediate-risk group, and 55.0% of the low-risk group. Median recipient age ranged from 42.5 (36.5-52.8) years in the low-risk cluster to 51.5 (39.5-62.8) years in the intermediate-risk cluster. Median donor age was similar across groups (32.5-35.5 years).

Immunologic parameters showed significant differences across clusters. Median HLA mismatches were highest in the high-risk group [8.0 (6.2-10.0)] and lowest in the intermediate-risk cluster [6.0 (5.0-7.0), P = 0.0000]. The number of DSAs also differed significantly, with medians of 5.0 (2.0-7.8), 3.0 (2.0-6.0), and 1.0 (1.0-2.2) in high-, intermediate-, and low-risk clusters, respectively (P = 0.0005). Similarly, cumulative DSA strength (MFI) was highest in the high-risk cluster [3811.0 (2366.5-6998.8)] and lowest in the low-risk cluster [686.5 (612.5-857.5), P = 0.0000].

Treatment characteristics and sensitizing events did not differ significantly between clusters, although IVIG use showed a borderline trend (high-risk: 69.0%, intermediate-risk: 55.9%, low-risk: 40.0%, P = 0.0637). Graft outcomes, including acute rejection, graft loss, DGF, and mortality, did not show statistically significant differences between clusters; however, rates were numerically higher in the high-risk cluster, with rates of acute rejection (6.9%), graft loss (6.9%), and mortality (5.2%) being notably higher. Median early graft function, assessed by creatinine and GFR at 4 months, was similar across groups.

Overall, univariate analysis confirms that the three clusters differ primarily in immunologic burden (HLA mismatches, DSA number, and cumulative MFI). At the same time, baseline demographics, treatments, and early clinical outcomes were broadly comparable. Table 4 shows the univariate analysis of various features of each cluster.

Table 4 Comparison of cluster features (univariate analysis), n (%).

Variable	High risk (n = 58)	Intermediate risk (n = 34)	Low risk (n = 20)	P value	Significance
Donor and recipient
Recipient gender				0.9388	Not significant
Male	30 (51.7)	17 (50.0)	11 (55.0)
Female	28 (48.3)	17 (50.0)	9 (45.0)
Recipient age, median (IQR)	46.5 (31.5-59.8)	51.5 (39.5-2.8)	42.5 (36.5-52.8)	0.1989	Not significant
Donor gender				0.2735	Not significant
Male	37 (63.8)	27 (79.4)	13 (65.0)
Female	21 (36.2)	7 (20.6)	7 (35.0)
Donor age, median (IQR)	35.5 (29.0-44.0)	35.0 (29.0-43.2)	32.5 (27.2-39.2)	0.3195	Not significant
Donor type				0.3791	Not significant
Living donor	52 (89.7)	33 (97.1)	19 (95.0)
Deceased donor	6 (10.3)	1 (2.9)	1 (5.0)
DSA and HLA matching
HLA mismatches, median (IQR)	8.0 (6.2-10.0)	6.0 (5.0-7.0)	7.0 (5.0-9.0)	0.0000	Significant
DSA type	Class I + II, 58 (100.0)	Class II only, 34 (100.0)	Class I, 19 (95.0), Class II, 1 (5.0)	-	Not tested
Number of DSAs, median (IQR)	5.0 (2.0-7.8)	3.0 (2.0-6.0)	1.0 (1.0-2.2)	0.0005	Significant
Cumulative DSA strength (MFI), median (IQR)	3811.0 (2366.5-6998.8)	2277.0 (1213.8-3900.8)	686.5 (612.5-857.5)	0.0000	Significant
Treatment and sensitization
Sensitizing events
Sensitized	5 (8.6)	3 (8.8)	1 (5.0)	1.0000	Not significant
Not sensitized	53 (91.4)	31 (91.2)	19 (95.0)
Induction therapy
ATG	58 (100.0)	31 (91.2)	19 (95.0)	0.0825	Borderline
Basiliximab	0 (0.0)	3 (8.8)	1 (5.0)
Maintenance therapy
FK MMF PRED	57 (98.3)	34 (100.0)	20 (100.0)	0.6252	Not significant
CSA MMF PRED	1 (1.7)	0 (0.0)	0 (0.0)
Diabetes pre-transplant
Yes	27 (46.6)	18 (52.9)	7 (35.0)	0.4426	Not significant
No	31 (53.4)	16 (47.1)	13 (65.0)
Previous transplant
Yes	2 (3.4)	0 (0.0)	1 (5.0)	-	Not tested
No	56 (96.6)	4 (100.0)	19 (95.0)
IVIG use
Yes	40 (69.0)	19 (55.9)	8 (40.0)	0.0637	Borderline
No	18 (31.0)	15 (44.1)	12 (60.0)
ATG dose	5.0 (4.0-5.0)	5.0 (4.0-5.0)	5.0 (4.0-5.0)	0.7762	Not significant
Graft function and outcomes
Acute rejection
Yes	4 (6.9)	0 (0.0)	1 (5.0)	0.3002	Not significant
No	54 (93.1)	20 (100.0)	19 (95.0)
Graft loss
Yes	4 (6.9)	2 (5.9)	0 (0.0)	0.4912	Not significant
No	54 (93.1)	32 (94.1)	20 (100.0)
Delayed graft function				0.5298	Not significant
Yes	2 (3.4)	2 (5.9)	0 (0.0)
No	56 (96.6)	32 (94.1)	20 (100.0)
Death (mortality)
Yes	3 (5.2)	2 (5.9)	0 (0.0)	0.5590	Not significant
No	55 (94.8)	32 (94.1)	20 (100.0)
CMV infection
Yes	2 (3.4)	1 (2.9)	0 (0.0)	0.7078	Not significant
No	56 (96.6)	19 (97.1)	20 (100.0)
BKV infection
Yes	1 (1.8)	1 (3.3)	0 (0.0)	-	Not tested
No	54 (98.2)	29 (96.7)	19 (100.0)
Length of stay (days)	9.0 (8.0-12.0)	10.0 (8.0-11.0)	9.0 (7.0-12.5)	0.9692	Not significant
Creatinine at 4 months (µmol/L)	94.0 (76.0-114.0)	88.0 (70.5-110.8)	77.5 (69.8-103.8)	0.4911	Not significant
eGFR at 4 months (mL/minute/1.73 m2)	76.0 (64.0-88.0)	78.0 (59.5-97.5)	92.0 (66.2-109.5)	0.3234	Not significant

P value < 0.05 was considered statistically significant. Continuous variables were analyzed using the Kruskal-Wallis test, and categorical variables were assessed using the χ² test as appropriate, based on expected cell counts. DSA: Donor-specific antibody; HLA: Human leukocyte antigen; ATG: Anti-thymocyte globulin; FK: Tacrolimus; MMF: Mycophenolate mofetil; CSA: Cyclosporine A; PRED: Prednisolone; IVIG: Intravenous immunoglobulin; eGFR: Estimated glomerular filtration rate; MFI: Mean fluorescence intensity; CMV: Cytomegalovirus; IQR: Interquartile range; BKV: BK virus.

Immunologic stratification and antibody patterns by risk cluster

Median immunologic and early clinical features varied distinctly across the three risk clusters. The high-risk cluster (n = 58) had the highest immunologic burden. The cluster included both Class I and Class II DSAs (MFI 1174 and 1966, respectively), with a median of 5 antibodies and 8 HLA mismatches. This group had relatively lower early graft function (GFR, 76 mL/minute/1.73 m²).

The intermediate-risk cluster (n = 34) was characterized by predominantly Class II sensitization (MFI 2277), a moderate DSA count of 3, fewer HLA mismatches (6), and slightly higher early graft function (GFR 78 mL/minute/1.73 m²).

In contrast, the low-risk cluster (n = 20) exhibited a minimal immunologic burden, characterized by isolated Class I DSA (MFI 686), a single antibody, 7 HLA mismatches, and the highest early graft function (GFR 92 mL/minute/1.73 m²). These findings highlight a precise gradation of immunologic risk across clusters, which corresponds with early graft performance. Table 5 presents the median values of key immunologic and clinical features for each risk cluster.

Table 5 Immunologic stratification and antibody patterns by risk cluster (median values).

Risk cluster	Class I MFI	Class II MFI	eGFR at 4 months (mL/minute/1.73 m2)	HLA mismatches	DSA type	DSA count
High risk cluster	1174	1966	76	8	Both	5
Intermediate risk cluster	0	2277	78	6	Class II	3
Low risk cluster	686	0	92	7	Class I	1

MFI: Mean fluorescence intensity; eGFR: Estimated glomerular filtration rate; HLA: Human leukocyte antigen; DSA: Donor-specific antibody.

Figure 2 illustrates a PCA. It is based on the visualization of the three immunologic risk clusters. Each point represents a DSA-positive KTR and is color-coded according to its assigned cluster. Shaded contours illustrate the distribution and density of patients within each cluster, highlighting areas where recipients are more closely grouped. The high-risk cluster (red) shows a tightly grouped and clearly separated pattern, representing patients with numerous DSAs, high MFI values, and sensitization to both Class I and Class II antigens. The intermediate-risk cluster (orange) exhibits partial separation and primarily comprises patients with Class II DSA and a moderate level of immunologic risk. The low-risk cluster (green) appears as a distinct and well-separated group, consisting of patients with isolated Class I DSA, low MFI values, and good early graft function. Overall, the PCA plot supports clear stratification and consistency within the clustering model.

Figure 2 illustrates a principal component analysis. Each point in this principal component analysis plot represents a donor-specific antibody (DSA)-positive kidney transplant recipient, color-coded by cluster membership, with shaded contours indicating the spatial density of patients within each group. The high-risk cluster (orange) demonstrates tight cohesion and distinct localization, characterized by patients with high mean fluorescence intensity (MFI), multiple DSAs, and mixed Class I and II sensitization. The intermediate-risk cluster (yellow) shows moderate density and spatial separation, predominantly representing Class II DSA patients with moderate immunologic burden. The low-risk cluster (green) occupies a well-isolated region with minimal overlap, encompassing patients with Class I-only DSA, low MFI, and favorable early graft function. While some overlap is observed between the high-risk and intermediate-risk clusters, the core density of each group remains distinct, indicating meaningful stratification. PCA: Principal component analysis.

To further characterize the immunologic features underlying each cluster, three UpSet plots (Figure 3) were generated to visualize the distribution and overlap of antibody specificities. The first plot includes patients from both the high- and intermediate-risk groups who experienced adverse outcomes (acute rejection, graft loss, or mortality). The most frequent antibody combinations include DR, DP, and DQ. Several unique multi-locus patterns were also observed, indicating considerable immunologic diversity among affected patients. The second plot represents the intermediate-risk group and shows a more limited range of antibody combinations. Most patients had antibodies against DQ alone or in combination with DR or A, indicating a narrower sensitization pattern primarily confined to Class II loci. The third plot, for the high-risk group, displays the greatest antibody complexity, with multiple frequent combinations involving DR, DP, and C. Several patients had three or more concurrent antibodies, indicating a broader level of sensitization. Together, these plots demonstrate that antibody diversity increases with immunologic risk, and patients with complex DSA profiles are more likely to experience adverse post-transplant outcomes.

Figure 3 UpSet plots illustrating antibody combination patterns across different risk groups and outcomes in kidney transplant recipients. A: Patients from high- and intermediate-risk groups with adverse outcomes (acute rejection, graft loss, or mortality); B: Intermediate-risk group; C: High-risk group. The most frequent antibody combinations involve DR, DP, and DQ, with multiple unique multi-locus patterns, reflecting immunologic diversity among affected individuals. HLA loci are indicated as A: HLA-A, B: HLA-B, C: HLA-C. PCA: Principal component analysis.

Clinical outcomes

Adverse clinical outcomes varied across the immunologic risk clusters, reflecting differences in underlying DSA profiles. The high-risk cluster experienced the highest rates of acute rejection and graft loss, consistent with its elevated immunologic burden, including mixed Class I and II DSA, higher MFI values, and greater HLA mismatches. The intermediate-risk cluster had no cases of acute rejection but showed modest rates of graft loss, DGF, and mortality. In contrast, the low-risk cluster exhibited the most favorable clinical course, with only a single case of acute rejection and no instances of graft loss or mortality (Table 6). Acute rejection episodes were further subclassified according to Banff criteria into T-cell–mediated rejection (TCMR), antibody-mediated rejection (ABMR), and mixed rejection. In the high-risk cluster, one episode was ABMR, one was TCMR, and two were mixed rejection. The single rejection episode in the low-risk cluster was TCMR.

Table 6 Frequency (and percentage) of key adverse clinical outcomes observed within each immunologic risk cluster, n (%).

Cluster	Acute rejection	Graft loss	Delayed graft function	Mortality
High risk cluster	4 (6.9)	4 (6.9)	2 (3.4)	3 (5.2)
Intermediate risk	0 (0.0)	2 (5.9)	2 (5.9)	2 (5.9)
Low risk cluster	1 (5.0)	0 (0.0)	0 (0.0)	0 (0.0)

These patterns were further supported by Kaplan–Meier analysis of event-free survival (Figure 4). The study was performed to compare event-free survival among DSA-negative, intermediate-risk, and high-risk KTR. The high-risk group showed a trend toward lower survival compared with the other groups. This reflects a higher likelihood of experiencing acute rejection, graft loss, or mortality. Although this difference did not reach statistical significance (log-rank P = 0.0849), the pattern suggests that immunologic risk may influence early post-transplant outcomes. Despite the lack of statistical significance on Kaplan–Meier analysis, the high-risk cluster consistently demonstrated numerically poorer event-free survival, suggesting an underlying adverse outcome trend.

Figure 4 Kaplan-Meier survival curve for composite outcome (acute rejection, graft loss, or mortality) by immunologic risk group. Kaplan-Meier analysis comparing composite event-free survival among donor-specific antibody -negative, high-risk, and intermediate-risk groups showed a non-significant but suggestive difference (log-rank P = 0.0849). Survival rates appeared lower in the High-Risk group, although this did not reach statistical significance. AR: Acute rejection; DSA: Donor-specific antibody.

DISCUSSION

To our knowledge, this is the first study from Saudi Arabia to apply an unsupervised machine learning approach for immunologic risk stratification DSA-positive KTR. Using K-means clustering, we identified three distinct risk clusters. These clusters showed clear differences in immunologic burden, antibody profiles, and early post-transplant outcomes. The high-risk cluster is characterized by broad Class I and II sensitization, higher DSA numbers, and stronger MFI values. This cluster demonstrated the greatest incidence of acute rejection, graft loss, and mortality. In contrast, the low-risk cluster comprises patients with isolated Class I DSA and low MFI. This cluster demonstrated excellent early graft function, with no graft loss or mortality. These findings highlight the heterogeneity within the DSA-positive recipient population and underscore the clinical utility of data-driven clustering for immunologic risk assessment in kidney transplantation.

Our study demonstrates that machine-learning–based risk stratification can identify KTR with differing immunologic risk. High-risk patients experienced more mixed and antibody-mediated rejections, consistent with their elevated immunologic burden, whereas low-risk patients predominantly had T-cell-mediated episodes. It is important to note the fragility of these results: Given the relatively small sample size, a single acute rejection event could substantially alter the observed risk distribution. Median cumulative DSA strength differed significantly across clusters [high-risk: 3811 (2366-6999); intermediate-risk: 2277 (1214-3901); low-risk: 686 (613-858); P < 0.001], reflecting a graded immunologic burden. Importantly, all recipients were included regardless of MFI if the crossmatch was negative, and all cluster medians remained below the conventional high-risk threshold of 5000 MFI. This supports the validity of our stratification while emphasizing that MFI values should not be interpreted in isolation. Instead, antibody specificity, complement-binding capacity, crossmatch results, and clinical context should inform graft allocation and post-transplant management. Overall, these observations underscore both the potential of machine learning tools to refine immunologic risk prediction and the need for cautious interpretation in small cohorts.

Previous studies from the Middle East have mainly focused on describing the prevalence of DSA, desensitization protocols, and outcomes of HLA-incompatible transplantation[13-15]. These reports lacked detailed immunologic phenotyping and long-term risk stratification. None of these studies has applied unsupervised clustering or similar machine-learning methods to characterize immunologic heterogeneity among DSA-positive recipients. Interest in unsupervised learning in transplantation, utilizing artificial intelligence in precision medicine, is growing[16]. Internationally, unsupervised and multi view clustering approaches have been used to define clinically meaningful transplant phenotypes. For example, a large semi supervised analysis of over 7000 kidney-transplant biopsies redefined Banff rejection categories into six data-driven phenotypes. Each phenotype showed stronger associations with graft failure than the traditional schema[9]. Another study in HLA-incompatible kidney transplantation used unsupervised clustering to analyze early post-transplant DSA responses. It identified five distinct patterns: No response, fast modulation, slow modulation, rise to sustained, and sustained. Modulation groups had higher rates of early acute rejection, but the lowest rates of five-year graft failure. On the other side, sustained responders had lower rates of early rejection but worse long-term outcomes, highlighting the prognostic value of early DSA dynamics[17]. Another study done in the United States used consensus clustering among Black KTR. It identified four phenotypes: Highly sensitized retransplants, low-risk living donor recipients, young hypertensive non-diabetic recipients, and older diabetic recipients with high-risk donors. These groups showed clear differences in graft and patient outcomes. This demonstrates the utility of unsupervised learning in revealing clinically meaningful subgroups[18]. Similarly, recipients with DGF were grouped into four clusters in a study. These four clusters included young sensitized retransplants, older diabetics with high kidney donor profile index kidneys, young non-diabetic black recipients, and middle-aged recipients with diabetes or hypertension. Although outcomes varied slightly, death-censored graft loss was similar, suggesting comorbidities may drive outcomes more than DGF itself[19]. Finally, in older recipients, unsupervised clustering of donor characteristics identified five donor phenotypes. Clusters 1 and 5 included younger, healthier donors, while clusters 2-4 comprised older donors with more comorbidities, with cluster 4 having the highest rates of hypertension, diabetes, and cerebrovascular death. Clusters 3 & 4 were associated with a higher risk of graft failure, even when kidney donor risk index scores were similar, highlighting the ability of machine learning to uncover hidden donor risk patterns[20]. Our study aligns with this global trend, demonstrating the feasibility of applying artificial intelligence-based clustering to identify biologically and clinically meaningful subgroups within a complex immunologic dataset.

Class I DSA generally disappears over time[21]. In contrast, class II DSA (especially DSA against DQ) persists for longer[22]. Class II DSA are less effectively reduced by desensitization therapies, such as plasmapheresis, compared to Class I DSA[23,24]. Previous studies have shown that higher DSA strength[4] and multiple antibody specificities, particularly against both Class I and II antigens[24], are associated with poorer allograft outcomes. Our results align with these observations. The high-risk cluster in our study included recipients with both Class I and Class II DSA. These patients had higher MFI levels and broader antibody reactivity. They experienced the worst graft outcomes, including more frequent rejection, graft loss, and death. This supports the concept that the combination of Class I and Class II DSA, primarily when associated with high antibody strength and diversity, confers a higher immunologic risk profile.

The distribution of median feature values across the three clusters further illustrates these distinctions. The high-risk cluster, defined by the coexistence of both Class I and II DSA, showed higher median MFI values (Class I: 1174; Class II: 1966) and a greater DSA count of 5. This reflects broad and intense antibody reactivity. The intermediate-risk cluster exhibited isolated Class II DSA with a median Class II MFI of 2277 and a DSA count of 3. This suggests persistent Class II dominant immune activation. In contrast, the low-risk cluster comprised recipients with isolated Class I DSA of low strength (median MFI: 686) and single antibody specificity. This represents a limited sensitization pattern. Consistent with these immunologic profiles, Kaplan-Meier analysis of the composite outcome (acute rejection, graft loss, or mortality) indicated a trend toward lower event-free survival in the high-risk cluster compared with intermediate-risk and DSA-negative recipients. However, the difference did not reach statistical significance (log-rank P = 0.0849). These observations align with data from the Swiss Transplant Cohort Study, which showed that pre-transplant Class II DSA, even at modest MFI levels, is associated with a higher risk of antibody-mediated rejection and graft loss, whereas isolated Class I DSA poses a lower risk unless the MFI is high[4]. Taken together, these findings support the utility of the clustering model in identifying patients with elevated immunologic risk.

Underlying immunologic mechanisms can explain the observed association between DSA characteristics and early post-transplant outcomes. A broader and stronger DSA profile, mainly when it includes both Class I and Class II antibodies, can amplify immune injury. This combined response may promote complement activation, endothelial damage, and microvascular inflammation, thereby increasing the risk of acute rejection and graft loss[25,26]. Class II DSA which tends to persist longer and is less effectively reduced by desensitization therapies. This may contribute to chronic antibody-mediated injury and progressive graft dysfunction over time[23,24]. Conversely, isolated low-strength DSA are less likely to elicit substantial immune activation, which aligns with favorable early graft outcomes[27]. These mechanistic considerations underscore the importance of both antibody strength and breadth in risk stratification of KTR.

Categorizing immunological risk in clusters has potential clinical utility. Identification of patients with higher DSA strength and broader antibody profiles could help guide personalized post-transplant monitoring. This may help in closer surveillance, earlier graft biopsies, or more intensive immunosuppression strategies in the high-risk group. Additionally, incorporating antibody characteristics such as class, number, and cumulative MFI into pre-transplant risk stratification or virtual cross-match algorithms may help identify recipients at higher immunologic risk. This may help optimize donor selection and guide desensitization approaches. Although early outcomes were not statistically different, the numerically poorer outcomes observed in the high-risk cluster suggest that these immunologic profiles may still be clinically meaningful for long-term risk assessment and individualized management.

This study has several notable strengths. It is the first from Saudi Arabia to apply unsupervised machine learning for immunologic risk stratification in DSA-positive KTR. Our study demonstrates the feasibility and utility of a data-driven approach. The cohort was largely homogeneous, consisting primarily of living-donor kidney transplants. This minimizes confounding from donor type, allowing for clearer interpretation of immunologic risk. A detailed assessment of the DSA class, number, cumulative MFI, and HLA mismatches, complemented by PCA and UpSet plot analyses, enabled a robust characterization of antibody diversity. Consistent induction and maintenance of immunosuppression further ensured the high quality of the data. Finally, the study offers biologically plausible mechanistic insights into the relationship between antibody burden and diversity, and potential graft injury, establishing a framework for future long-term and multicenter validation studies.

This study has several limitations. The sample size was relatively small. In addition, there were a few adverse events, which limited the statistical power to detect differences in early clinical outcomes between clusters. The cohort was largely homogeneous, mainly consisting of living-donor transplants, which may limit generalizability to deceased-donor populations or centers with more diverse transplant practices. The follow-up period was relatively short, preventing assessment of long-term graft outcomes and the potential delayed impact of persistent Class II DSA. Although clustering highlighted immunologic patterns, other factors that influence graft outcomes, such as non-HLA antibodies, adherence, or subclinical inflammation, were not evaluated. As with any single-center retrospective study, unmeasured confounders may also have affected both immunologic profiles and clinical outcomes.

Future larger studies are needed to validate these immunologic clusters in larger, multicenter cohorts consisting of both live and deceased donors. A longer follow-up is required to assess the impact of persistent Class II DSA and antibody diversity on long-term graft survival and chronic rejection. Integrating other immunologic markers, such as non-HLA antibodies, complement activation profiles, or molecular injury signatures, could enhance risk stratification. Ultimately, prospective studies could investigate whether tailored monitoring, early biopsy, or individualized immunosuppression guided by cluster-based risk profiles enhances graft outcomes in DSA-positive KTR.

CONCLUSION

Unsupervised machine learning can identify distinct immunologic risk groups among DSA-positive KTR. Patients with both Class I and Class II DSA, higher MFI levels, and broader antibody patterns formed a high-risk cluster, characterized by the poorest early outcomes. In contrast, isolated low-strength Class I DSA was associated with excellent short-term graft function. These findings highlight the heterogeneity within the DSA-positive population and the value of clustering for refined risk stratification. Larger multicenter studies with longer follow-up are needed to validate these clusters and determine their role in guiding personalized post-transplant care.