Keywords: cadaver organ transplantation, chronic allograft failure, kidney transplantation, mortality, mortality risk, renal transplantation, transplant outcomes, transplantation
Abstract
Key Points
An unsupervised machine learning clustering algorithm identified distinct deceased kidney donor phenotypes among older recipients.
Recipients of certain donor phenotypes were at a relatively higher risk of all-cause graft loss even after accounting for recipient factors.
The use of unsupervised clustering to support kidney allocation systems may be an important area for future study.
Background
Older transplant recipients are at a relatively increased risk of graft failure after transplantation, and some of this risk may relate to donor characteristics. Unsupervised clustering using machine learning may be a novel approach to identify donor phenotypes that may then be used to evaluate outcomes for older recipients. Using a cohort of older recipients, the purpose of this study was to (1) use unsupervised clustering to identify donor phenotypes and (2) determine the risk of death/graft failure for recipients of each donor phenotype.
Methods
We analyzed a nationally representative cohort of kidney transplant recipients aged 65 years or older captured using the Scientific Registry of Transplant Recipients between 2000 and 2017. Unsupervised clustering was used to generate phenotypes using donor characteristics inclusive of variables in the kidney donor risk index (KDRI). Cluster assignment was internally validated. Outcomes included all-cause graft failure (including mortality) and delayed graft function. Differences in the distribution of KDRI scores were also compared across the clusters. All-cause graft failure was compared for recipients of donor kidneys from each cluster using a multivariable Cox survival analysis.
Results
Overall, 23,558 donors were separated into five clusters. The area under the curve for internal validation of cluster assignment was 0.89. Recipients of donor kidneys from two clusters were found to be at high risk of all-cause graft failure relative to the lowest risk cluster (adjusted hazards ratio, 1.86; 95% confidence interval, 1.69 to 2.05 and 1.73; 95% confidence interval, 1.61 to 1.87). Only one of these high-risk clusters had high proportions of donors with established risk factors (i.e., hypertension, diabetes). KDRI scores were similar for the highest and lowest risk clusters (1.40 [1.18–1.67] and 1.37 [1.15–1.65], respectively).
Conclusions
Unsupervised clustering can identify novel donor phenotypes comprising established donor characteristics that, in turn, may be associated with different risks of graft loss for older transplant recipients.
Introduction
Kidney transplantation is beneficial for older recipients.1 Nevertheless, older adults are at an increased risk of death after transplantation,2–5 and organs from higher-risk deceased donors are associated with both death and graft failure in this population.6 However, while donor characteristics may have a direct effect on outcomes for older recipients, there are a paucity of approaches that currently exist to characterize donor phenotypes specifically for older adults.
One established approach to characterizing donor phenotypes is the kidney donor risk index (KDRI) which involves combining deceased donor characteristics to determine organ quality, which in turn, may be associated with outcomes after transplantation to inform organ allocation. Higher KDRI scores reflect donor kidneys of lower quality7 and hence, higher recipient risk for graft loss.8–11 While valuable, the KDRI has reduced utility for risk prediction for older recipients.12,13 Therefore, complementary strategies to capture combinations of donor factors unique to older recipients may be of value.
Phenotyping identifies groups of patients with similar characteristics allowing the administration of personalized interventions.14–17 Kidney donor phenotyping has been deemed to be of value for kidney allocation; identifying high-risk donors that may affect recipient outcomes is of national importance to ensure the objectives of kidney allocation which are to preserve donor kidney longevity and recipient life expectancy, while maintaining equity for transplantation.18–20
Machine learning (ML) can identify unique patterns within large datasets.21,22 Only a few studies have explored unsupervised ML methods, such as clustering algorithms that partition unlabeled datasets comprising multiple features into distinct phenotypes on basis of the inherent similarity between the data points. Consensus clustering has been used to identify phenotypes among Black kidney transplant recipients23 and in a small dataset, older transplant recipients.24 However, these analyses included donor, recipient, and surgical factors limiting the ability to specifically identify independent donor phenotypes associated with poor outcomes after transplantation.
In this study, we generated phenotypes of kidney donors using an unsupervised ML clustering method. Our primary objective was to identify and characterize phenotypes of established donor predictors. Secondarily, we sought to determine whether clusters were associated with differential risks of death/graft failure and delayed graft function (DGF) for older adults.
Methods
Data from the Scientific Registry of Transplant Recipients (SRTR) were used to identify all patients aged 65 years or older who underwent kidney transplantation from 2000 to 2017. The SRTR data system includes data on all donors, waitlisted candidates, and transplant recipients in the United States, submitted by the members of the Organ Procurement and Transplant Network. The Health Resources and Services Administration, US Department of Health and Human Services, provides oversight to the activities of the Organ Procurement and Transplant Network and SRTR contractors. The clinical and research activities being reported are consistent with the Principles of the Declaration of Istanbul as outlined in the Declaration of Istanbul on Organ Trafficking and Transplant Tourism. We included individuals with complete data across donor characteristics of interest and excluded patients receiving repeat or multiorgan transplants. This study was approved through the Nova Scotia Research Ethics Board.
Covariates/Confounders
To maximize utility in informing provider decision making at the time of donor offer, we used several donor variables known to be associated with graft failure to derive donor clusters. These included KDRI factors of age, height, weight, race (classified as Black, Other, or White on the basis of SRTR database documentation), hypertension, diabetes, cerebrovascular cause of death (COD), serum creatinine, hepatitis C status, and donation after circulatory death status8; donor sex and donor body mass index (BMI)25; use of vasodilators, vasopressin, or inotropic agents26; and donor cytomegalovirus status.27 We also reported on variables that were not included in the cluster derivation, including the KDRI,8 expanded criteria donor status, recipient age, sex, race, height, weight, BMI, Karnofsky Performance Score, cause of kidney failure, coronary artery disease, hypertension, peripheral vascular disease, malignancy, dialysis vintage, peak panel reactive antibody level, cold ischemia time, the number of human leukocyte antigen mismatches, and era (categorized as transplant year 2000–2005, 2006–2011, and 2012–2017).
Outcomes
The primary outcome was all-cause graft failure, defined as either death or graft failure. Graft failure was defined as a permanent return to dialysis or preemptive retransplantation. Secondary outcomes included death-censored graft failure, death with graft function, and DGF (dialysis within the first week post-transplant).
ML Framework for Generating Donor Phenotypes
A detailed description of the approach to data clustering is noted in Supplemental File 1. To summarize, ML-based clustering methods partition unlabeled data into distinct clusters in an unsupervised manner, whereby data grouped in a cluster are similar to each other. Our ML donor phenotyping framework comprised three main modules (Figure 1). The data embedding module was to reduce the high dimensionality of the original data, while maintaining its inherent properties, to render the data suitable for clustering. We used split autoencoders—a multiview variant of autoencoders28 that in an unbiased manner generates a lower dimensional dataset. The split autoencoders were trained with different parameters, and we selected the optimal reduced feature set on the basis of least testing error. The donor phenotyping module was responsible for data clustering to generate the donor phenotypes. We used a cotraining clustering approach to cluster the unlabeled multiview data. Both numerical and categorical features were separately analyzed by clustering algorithms, and the clusters were integrated using a consensus mechanism to yield the final clusters. Our multiview clustering approach29,30 ensured no information loss during cluster formation as is typical of single-view clustering methods. Given the multiview data, we experimented with both spectral (which generates nonconvex clusters) and partitioning (which generates convex shaped clusters) methods to achieve the most optimal clustering outcome. We used multiview variants of partitioning clustering methods (K-means and K-medoids on the basis of co-expectation–maximization algorithm)31 and graph clustering methods32 to form consensus-driven clusters. Clustering performance was measured using standard metrics of the silhouette and Davies Bouldin score which measure the intracluster data similarity to determine cluster quality.33 The final donor classification module was used to internally validate the donor phenotypes as derived by the clustering method34,35 by learning the donor phenotype associated with each cluster to its cluster label. Using 80% of training data from each cluster with its cluster label, we trained a multiclass stacked ensemble classifier, comprising XGBoost, neural network, logistic regression, and random forest algorithms, to predict the cluster label for unseen donor data. Parameter tuning and regularization were applied to avoid overfitting, while cross-validation was used during training to ensure generalizability beyond the given training data. Together with the statistical analysis (see below) identifying donor phenotypes for each cluster, the classification performance of the ensemble classifier internally validates the cluster-based phenotypes by mapping the donor characteristics for each cluster with the associated graft outcome as noted in the original data (note that outcomes were not used for clustering).
Figure 1.
Machine learning framework for donor phenotyping and outcome prediction. Starting with donor data, the framework performs data's dimensionality reduction using autoencoders and donor phenotyping through multiview clustering methods. Finally, classification methods are used to learn the mapping between data points in a cluster to their cluster label because this can be used to predict the phenotype of new donors.
Statistical Analysis
Baseline characteristics of each cluster were reported using univariable statistics. Differences across clusters were assessed using the chi-squared test for categorical variables and the Kruskal–Wallis test for continuous variables. Time to all-cause graft failure was compared using the Kaplan–Meier product-limit method. The association of cluster assignment with graft failure was determined using a multivariable Cox survival analysis adjusted for recipient/immunologic factors (as noted in the covariates section) and reported using relative hazards ratio (HR) with 95% confidence intervals (CIs). Proportionality was determined using visual examination of log-log survival plots. In the adjusted analysis, missing recipient data were handled by multiple imputation using chained equations.36 For the imputation, we included individual donor variables that were used to develop the donor clusters (Covariates, above), complete recipient variables, and the outcome of all-cause graft failure. The total number of generated imputed datasets was targeted to be approximately equal to the proportion of incomplete cases (i.e. 40 datasets). In secondary analyses, the association between cluster assignment and outcomes was repeated for death with graft function, death-censored graft failure, 1-year and 3-year all-cause graft failure, and DGF. The latter was analyzed using an adjusted logistic regression model and reported using odds ratios. In sensitivity analyses, we analyzed the association between cluster assignment and the primary outcome inclusive of era of transplant (acknowledging potential differences in approaches to transplantation over the study period) and using case-wise deletion instead of multiple imputation. To test for interactions, the primary outcome was assessed across subgroups of recipient diabetes status, sex, race (White versus Other), peak panel reactive antibody category (>80 versus ≤80),32 recipient age (using a threshold of ≥75), and high-risk KDRI threshold (KDRI ≥1.45 versus KDRI <1.45).8 Interactions were deemed significant using a P-value of< 0.05. All statistical analyses were performed using Stata, version 17, StataCorp.
Results
From 2000 to 2017, 165,090 recipients received a deceased donor kidney transplant. After exclusions, our dataset included 23,558 unique, first, kidney transplants for recipients aged 65 years or older (Figure 2). Characteristics and the proportion of each variable with missing data are noted in Supplemental Table 1.
Figure 2.
Flowchart of the breakdown of included participants.
The optimal number of clusters was found to be five. The multiview k-medoids method produced the best clustering performance (silhouette=0.85, Davies Bouldin=0.48). Figure 3 shows the distribution of data, labeled as cluster 1–5. Cluster-specific chord diagrams of individual variables within each cluster are noted in Figure 4. Baseline characteristics are noted in Table 1. Cluster 1 and 5 had the youngest mean donor ages (38 and 35 years), smallest proportion of female donors (32% and 34%), and low proportions of hypertension, diabetes, and donors with a cerebrovascular COD. The two differentiators between clusters 1 and 5 were donor obesity (BMI ≥35 in 22% versus 2%, respectively) and Black race (32% versus 3%, respectively). Clusters 2, 3, and 4 exhibited similar mean donor age (50, 50, and 53 years, respectively) and high proportions of female donors. Cluster 4 had the highest rates of donor hypertension (84%), cerebrovascular COD (78%), and donor diabetes (22%). Clusters 2, 3, and 4 differed in the proportion of Black donors and donor weight; clusters 2 and 3 had fewer Black donors (9% and 4%, respectively) and mean donor weights of 86 and 76 kg, respectively, while cluster 4 had a higher proportion of Black donors (22%) and highest mean weight (90 kg).
Figure 3.
Clustering results. t-SNE mapping depicts five distinct clusters of donors of organs allocated to older transplant recipients.
Figure 4.
Chord diagrams depicting key donor characteristics for each cluster, thus representing the influence of specific donor phenotypes.
Table 1.
Baseline characteristics by donor cluster
| Variable | Cluster 1 (N=1784) | Cluster 2 (N=1832) | Cluster 3 (N=2522) | Cluster 4 (N=9071) | Cluster 5 (N=8349) |
|---|---|---|---|---|---|
| Features included in cluster derivation | |||||
| Age mean±SDa | 38±15 | 50±14 | 50±14 | 53±11 | 35±16 |
| Female sexa | 565 (32) | 896 (49) | 1242 (49) | 4474 (49) | 2868 (34) |
| Donor racea,b | |||||
| Black | 575 (32) | 166 (9) | 100 (4) | 1963 (22) | 231 (3) |
| Other | 56 (3) | 60 (3) | 33 (1) | 467 (5) | 208 (2) |
| White | 1153 (65) | 1606 (88) | 2389 (95) | 6641 (73) | 7910 (95) |
| Donor heighta | 172±12 | 170±11 | 170±11 | 170±10 | 170±17 |
| Donor weighta | 87±23 | 86±21 | 76±17 | 90±23 | 73±20 |
| Donor BMI categoriesa | |||||
| <18.5 | 21 (1) | 34 (2) | 29 (1) | 68 (1) | 761 (9) |
| ≥18.5 to <25 | 521 (29) | 367 (20) | 1131 (45) | 1962 (22) | 3534 (42) |
| ≥25 to <30 | 536 (30) | 492 (27) | 1062 (42) | 2645 (29) | 2907 (35) |
| ≥30 to <34 | 319 (18) | 676 (37) | 206 (8) | 2043 (23) | 1014 (12) |
| ≥35 | 387 (22) | 263 (14) | 94 (4) | 2353 (26) | 133 (2) |
| Donor CMV positivea | 1199 (67) | 1317 (72) | 1333 (53) | 6228 (69) | 5006 (60) |
| Donor hypertensiona | 237 (13) | 621 (34) | 639 (25) | 7639 (84) | 71 (1) |
| Donor diabetesa | 90 (5) | 127 (7) | 58 (2) | 1959 (22) | 149 (2) |
| Donor hepatitis Ca | 27 (2) | 33 (2) | 44 (2) | 100 (2) | 325 (4) |
| Donor cerebrovascular CODa | 118 (7) | 900 (49) | 1634 (65) | 7091 (78) | 117 (1) |
| Donation after cardiac deatha | 376 (21) | 277 (15) | 166 (7) | 1373 (15) | 1252 (15) |
| Donor creatininea | 1.1 (0.8–1.5) | 1.0 (0.7–1.3) | 0.9 (0.7–1.2) | 1.0 (0.8–1.4) | 1.0 (0.7–1.3) |
| Vasodilator usea | 176 (10) | 213 (12) | 379 (15) | 1235 (14) | 1032 (12) |
| Vasopressin usea | 984 (55) | 990 (54) | 1483 (59) | 4617 (51) | 4791 (57) |
| Inotrope use | 857 (48) | 986 (54) | 1322 (52) | 4496 (50) | 4011 (48) |
| Features NOT included in cluster derivation | |||||
| Recipient characteristic | |||||
| Age mean±SDa | 69±4 | 69±4 | 69±4 | 70±4 | 69±4 |
| Female sexc | 664 (37) | 680 (37) | 989 (39) | 3308 (36) | 3236 (39) |
| Recipient racea,b | |||||
| Black | 453 (25) | 403 (22) | 542 (21) | 2324 (26) | 1772 (21) |
| Other | 126 (7) | 157 (9) | 188 (7) | 844 (9) | 663 (8) |
| White | 1205 (68) | 1272 (69) | 1792 (71) | 5903 (65) | 5914 (71) |
| Recipient heightc | 170±11 | 169±11 | 169±10 | 169±11 | 170±11 |
| Recipient weight | 82±18 | 80±17 | 80±16 | 81±17 | 81±17 |
| Recipient BMI categories | |||||
| <18.5 | 32 (2) | 25 (1) | 24 (1) | 118 (1) | 122 (2) |
| ≥18.5 to <25 | 430 (25) | 482 (27) | 659 (27) | 2290 (26) | 2216 (27) |
| ≥25 to <30 | 650 (38) | 669 (38) | 952 (39) | 3439 (40) | 3056 (38) |
| ≥30 to <34 | 444 (26) | 457 (26) | 585 (24) | 2066 (24) | 1971 (24) |
| ≥35 | 173 (10) | 137 (8) | 199 (8) | 762 (9) | 729 (9) |
| Requiring assistancea | 294 (17) | 228 (13) | 364 (15) | 1257 (14) | 1331 (17) |
| Cause of kidney failurea | |||||
| Diabetes | 692 (40) | 692 (39) | 931 (38) | 3534 (40) | 3100 (38) |
| Glomerulonephritis | 187 (11) | 199 (11) | 310 (13) | 995 (11) | 1007 (12) |
| Hypertension | 512 (29) | 546 (31) | 674 (28) | 2770 (31) | 2253 (28) |
| Other | 213 (12) | 219 (12) | 339 (14) | 961 (11) | 1169 (14) |
| Polycystic kidney disease | 134 (8) | 132 (7) | 194 (8) | 608 (7) | 605 (7) |
| Coronary artery disease | 182 (13) | 212 (15) | 226 (12) | 938 (13) | 861 (13) |
| Hypertension | 1396 (91) | 1464 (91) | 1957 (90) | 7280 (92) | 6470 (91) |
| Diabetesc | 896 (50) | 893 (49) | 1199 (48) | 4549 (50) | 3997 (48) |
| Cerebrovascular disease | 62 (4) | 71 (5) | 66 (3) | 297 (4) | 273 (4) |
| Peripheral vascular disease | 161 (9) | 172 (10) | 215 (9) | 821 (9) | 710 (9) |
| Malignancy | 242 (14) | 224 (12) | 336 (14) | 1230 (14) | 1179 (14) |
| Dialysis vintagea | 2.72 [0–4.95] | 2.88 [0.60–4.73] | 2.65 [0–4.84] | 2.84 [0.34–4.87] | 2.49 [0–4.68] |
| Peak PRA groupa | |||||
| <20 | 983 (73) | 1096 (74) | 1501 (76) | 5765 (79) | 4644 (72) |
| 20–80 | 250 (19) | 278 (19) | 348 (18) | 1159 (16) | 1266 (20) |
| >80 | 118 (9) | 113 (8) | 132 (7) | 338 (5) | 569 (9) |
| Other characteristics | |||||
| Cold ischemia timea | 16.5 (11–22.2) | 17 (11.9–23) | 15.85 (10.75–22.23) | 17.30 (12–24) | 16.13 (11–22.21) |
| HLA mismatchesa | |||||
| 0 | 91 (5) | 123 (7) | 228 (9) | 424 (5) | 869 (10) |
| 1 | 36 (2) | 25 (1) | 49 (2) | 81 (1) | 190 (2) |
| 2 | 73 (4) | 81 (4) | 94 (4) | 264 (3) | 332 (4) |
| 3 | 272 (15) | 247 (14) | 294 (12) | 1052 (12) | 1054 (13) |
| 4 | 465 (26) | 480 (26) | 674 (27) | 2321 (26) | 2143 (26) |
| 5 | 554 (31) | 576 (32) | 746 (30) | 3152 (35) | 2516 (30) |
| 6 | 270 (15) | 291 (16) | 418 (17) | 1735 (19) | 1178 (14) |
| Eraa | |||||
| 2000–2005 | 127 (7) | 182 (10) | 214 (8) | 842 (9) | 753 (9) |
| 2006–2011 | 737 (41) | 847 (46) | 1060 (42) | 4093 (45) | 3553 (43) |
| 2012–2017 | 920 (52) | 803 (44) | 1248 (49) | 4136 (46) | 4043 (48) |
BMI, body mass index; CMV, cytomegalovirus; COD, cause of death; PRA, peak panel reactive antibody; HLA, human leukocyte antigen.
P < 0.001.
The category other for race included SRTR race categories as noted in the SRTR data dictionary (https://www.srtr.org/requesting-srtr-data/saf-data-dictionary/).
P < 0.05.
The distribution of KDRI scores are noted in Table 2. Donors in cluster 4 had the highest median KDRI score (1.68, interquartile range [IQR] 1.44–1.97), while those in cluster 5 had the lowest (median KDRI 1.02, IQR 0.99–1.20).
Table 2.
Kidney Donor Risk Index Scores and Expanded Criteria Donor Status across each cluster
| Variable | Cluster 1 (N=1784) | Cluster 2 (N=1832) | Cluster 3 (N=2522) | Cluster 4 (N=9071) | Cluster 5 (N=8349) |
|---|---|---|---|---|---|
| ECDa | 190 (11) | 596 (33) | 806 (32) | 5282 (58) | 690 (8) |
| KDRIa | 1.12 (0.97–1.34) | 1.40 (1.18–1.67) | 1.37 (1.15–1.65) | 1.68 (1.44–1.97) | 1.02 (0.99–1.20) |
| High-risk KDRI group (≥1.45)a | 315 (18) | 826 (45) | 1061 (42) | 6704 (68) | 976 (12) |
ECD, expanded criteria donor; KDRI, kidney donor risk index.
P < 0.001.
Over 91,865 total years at risk (median follow-up time of 3.2 years, IQR 1.2–5.9), there were 9555 primary events. Recipients of kidneys from clusters 2 and 4 had the highest rates of death and graft failure (52% and 48%) as well as DGF (Table 2 and Supplemental Figure 1). One-year graft failure was highest for recipients of kidneys from clusters 2 and 4 (Supplemental Figure 1 and Supplemental Table 2). Time to all-cause graft failure is noted in Figure 5 (log rank P < 0.001). Recipients of kidneys from cluster 2 had the highest adjusted risk (adjusted hazards ratio, 1.86; 95% CI, 1.69 to 2.05), compared with cluster 3 (Table 3); recipients of kidneys from cluster 4 were also at high risk (adjusted hazards ratio, 1.73; 95% CI, 1.61 to 1.87; Table 3). A similar association was observed for 1-year and 3-year all-cause graft loss (Supplemental Tables 2 and 3). The risk of death with graft function was higher for recipients of cluster 2 kidneys and higher for death-censored graft failure for recipients of cluster 4 kidneys (Table 4). The risk of DGF was highest for cluster 4 (odds ratio, 2.02; 95% CI, 1.74 to 2.34). The results in the sensitivity analyses were consistent with the primary analysis. When looking across recipient characteristics (Table 5), recipient age was an effect modifier for recipients of kidneys from cluster 1 and cluster 5 (lower risk of all-cause graft failure in recipients aged 75 years or older). Female recipients of kidneys from cluster 4 were also at an increased risk of all-cause graft loss (relative to male recipients).
Figure 5.
Time to all-cause graft loss, stratified by each donor cluster group (log rank P < 0.001).
Table 3.
Time to all-cause graft failure in unadjusted and adjusted Cox proportional hazards models (N=23,558 individuals, 9555 events)
| Model (N/Events) | Relative Hazard | 95% CI |
|---|---|---|
| Unadjusted | ||
| Cluster 1 | 1.45 | 1.31 to 1.61 |
| Cluster 2 | 1.89 | 1.72 to 2.08 |
| Cluster 3 | Reference | — |
| Cluster 4 | 1.79 | 1.66 to 1.93 |
| Cluster 5 | 1.16 | 1.07 to 1.26 |
| Adjusted with multiple imputation | ||
| Cluster 1 | 1.41 | 1.28 to 1.57 |
| Cluster 2 | 1.86 | 1.69 to 2.05 |
| Cluster 3 | Reference | — |
| Cluster 4 | 1.73 | 1.61 to 1.87 |
| Cluster 5 | 1.17 | 1.08 to 1.27 |
| Cold ischemia time (each hour) | 1.01 | 1.00 to 1.01 |
| Recipient BMI | ||
| <18.5 | 1.35 | 1.12 to 1.62 |
| ≥18.5 to <25 | Reference | |
| ≥25 to <30 | 1.01 | 0.93 to 1.08 |
| ≥30 to <34 | 1.01 | 0.90 to 1.15 |
| ≥35 | 1.07 | 0.89 to 1.29 |
| History of malignancy | 0.92 | 0.87 to 0.98 |
| Coronary artery disease | 1.23 | 1.16 to 1.31 |
| Peripheral vascular disease | 1.23 | 1.14 to 1.32 |
| Cerebrovascular disease | 1.06 | 0.95 to 1.18 |
| Diabetes | 1.28 | 1.20 to 1.37 |
| Diabetes as cause of kidney failure | 1.07 | 1.00 to 1.15 |
| Male sex | 1.10 | 1.04 to 1.16 |
| Race | ||
| Black | Reference | — |
| Other | 0.85 | 0.77 to 0.93 |
| White | 1.05 | 1.00 to 1.10 |
| Low functional status (Karnofsky score ≤60) | 1.26 | 1.17 to 1.36 |
| Number of HLA mismatches | ||
| 0 | Reference | — |
| 1 | 1.10 | 0.93 to 1.29 |
| 2 | 1.03 | 0.90 to 1.17 |
| 3 | 1.08 | 0.99 to 1.19 |
| 4 | 1.08 | 0.99 to 1.17 |
| 5 | 1.10 | 1.01 to 1.19 |
| 6 | 1.13 | 1.04 to 1.24 |
| Peak PRA group | ||
| <20 | 0.99 | 0.93 to 1.04 |
| 20–80 | Reference | — |
| >80 | 1.10 | 1.00 to 1.22 |
| Dialysis vintage (each year) | 1.04 | 1.03 to 1.04 |
| Recipient hypertension (each cm) | 1.00 | 1.00 to 1.01 |
| Recipient weight (each kg) | 1.00 | 1.00 to 1.01 |
| Recipient age (each year) | 1.04 | 1.04 to 1.05 |
CI, confidence interval; BMI, body mass index; HLA, human leukocyte antigen; PRA, peak panel reactive antibody.
Table 4.
Secondary outcomes and additional sensitivity analyses
| Model (N/Events) | Relative Hazard | 95% CI |
|---|---|---|
| Adjusted primary analysis inclusive of era of transplant in multivariable model (23558/9555) | ||
| Cluster 1 | 1.42 | 1.28 to 1.57 |
| Cluster 2 | 1.85 | 1.68 to 2.03 |
| Cluster 3 | Reference | — |
| Cluster 4 | 1.72 | 1.60 to 1.86 |
| Cluster 5 | 1.17 | 1.08 to 1.26 |
| Adjusted primary analysis using case-wise deletion (13246/6281) | ||
| Cluster 1 | 1.49 | 1.31 to 1.69 |
| Cluster 2 | 1.96 | 1.75 to 2.20 |
| Cluster 3 | Reference | — |
| Cluster 4 | 1.77 | 1.61 to 1.95 |
| Cluster 5 | 1.20 | 1.09 to 1.33 |
| Death with graft function (23558/6958) | ||
| Cluster 1 | 1.48 | 1.32 to 1.67 |
| Cluster 2 | 1.93 | 1.73 to 2.16 |
| Cluster 3 | Reference | — |
| Cluster 4 | 1.62 | 1.48 to 1.77 |
| Cluster 5 | 1.29 | 1.18 to 1.42 |
| Death-censored graft failure (23558/2597) | ||
| Cluster 1 | 1.24 | 1.01 to 1.52 |
| Cluster 2 | 1.70 | 1.41 to 2.05 |
| Cluster 3 | Reference | — |
| Cluster 4 | 2.02 | 1.74 to 2.34 |
| Cluster 5 | 0.84 | 0.72 to 0.99 |
| DGF | ||
| Cluster 1 | 1.24 | 1.01 to 1.52 |
| Cluster 2 | 1.70 | 1.41 to 2.05 |
| Cluster 3 | Reference | — |
| Cluster 4 | 2.02 | 1.74 to 2.34 |
| Cluster 5 | 0.84 | 0.71 to 0.99 |
All models adjusted for cold ischemia time, recipient factors (body mass index, cause of kidney failure, race, age, functional status, dialysis vintage, height, weight, peak panel reactive antibody status, malignancy, coronary artery disease, peripheral vascular disease, cardiovascular disease, and diabetes), and number of human leukocyte antigen mismatches unless otherwise specified. CI, confidence interval; DGF, delayed graft function.
Table 5.
Time to all-cause graft loss across recipient strata
| Modela | Relative Hazard (95% CI) | Relative Hazard (95% CI) | Interaction P |
|---|---|---|---|
| Age | 75 yr or older | 75 yr or younger | |
| Cluster 1 | 2.75 (1.76 to 4.32) | 1.41 (1.23 to 1.61) | <0.001 |
| Cluster 2 | 2.23 (1.46 to 3.41) | 1.93 (1.71 to 2.18) | 0.45 |
| Cluster 3 | Reference | Reference | — |
| Cluster 4 | 2.09 (1.49 to 2.94) | 1.75 (1.59 to 1.94) | 0.36 |
| Cluster 5 | 1.62 (1.14 to 2.31) | 1.15 (1.04 to 1.28) | 0.02 |
| Diabetes | Yes | No | |
| Cluster 1 | 1.35 (1.17 to 1.56) | 1.47 (1.27 to 1.72) | 0.36 |
| Cluster 2 | 1.76 (1.54 to 2.01) | 1.99 (1.73 to 2.29) | 0.18 |
| Cluster 3 | Reference | Reference | — |
| Cluster 4 | 1.73 (1.55 to 1.92) | 1.78 (1.59 to 2.00) | 0.48 |
| Cluster 5 | 1.13 (1.01 to 1.26) | 1.20 (1.07 to 1.35) | 0.52 |
| Panel reactive antibody | >80 | ≤80 | |
| Cluster 1 | 1.57 (1.05 to 2.33) | 1.42 (1.27 to 1.59) | 0.73 |
| Cluster 2 | 2.00 (1.37 to 2.91) | 1.87 (1.69 to 2.07) | 0.76 |
| Cluster 3 | Reference | Reference | — |
| Cluster 4 | 1.95 (1.41 to 2.70) | 1.75 (1.61 to 1.91) | 0.45 |
| Cluster 5 | 1.08 (0.79 to 1.49) | 1.18 (1.08 to 1.28) | 0.46 |
| Sex | Male | Female | |
| Cluster 1 | 1.35 (1.19 to 1.53) | 1.51 (1.26 to 1.80) | 0.29 |
| Cluster 2 | 1.80 (1.60 to 2.02) | 2.00 (1.70 to 2.36) | 0.24 |
| Cluster 3 | Reference | Reference | — |
| Cluster 4 | 1.66 (1.51 to 1.82) | 1.95 (1.71 to 2.23) | 0.04 |
| Cluster 5 | 1.11 (1.01 to 1.22) | 1.24 (1.08 to 1.43) | 0.22 |
| Race | White | Other | |
| Cluster 1 | 1.36 (1.20 to 1.54) | 1.52 (1.25 to 1.85) | 0.72 |
| Cluster 2 | 1.87 (1.68 to 2.10) | 1.83 (1.52 to 2.20) | 0.82 |
| Cluster 3 | Reference | Reference | — |
| Cluster 4 | 1.72 (1.57 to 1.88) | 1.82 (1.57 to 2.12) | 0.79 |
| Cluster 5 | 1.17 (1.06 to 1.28) | 1.11 (0.95 to 1.30) | 0.88 |
| KDRI | Lower risk (<1.45) | High risk (≥1.45) | |
| Cluster 1 | 1.45 (1.28 to 1.65) | 1.59 (1.30 to 1.94) | 0.86 |
| Cluster 2 | 1.73 (1.51 to 1.97) | 2.00 (1.74 to 2.29) | 0.95 |
| Cluster 3 | Reference | Reference | — |
| Cluster 4 | 1.59 (1.42 to 1.78) | 1.68 (1.50 to 1.88) | 0.91 |
| Cluster 5 | 1.20 (1.08 to 1.32) | 1.33 (1.15 to 1.54) | 0.93 |
CI, confidence interval; KDRI, kidney donor risk index.
Adjusted for cold ischemia time, recipient factors (body mass index, cause of kidney failure, race, age, functional status, dialysis vintage, height, weight, peak panel reactive antibody status, malignancy, coronary artery disease, peripheral vascular disease, cardiovascular disease, diabetes), and number of human leukocyte antigen mismatches.
Internal validation of the cluster-specific donor phenotypes to their respective cluster label yielded an F1 score of 0.88. Supplemental Table 4 shows the classification performance of the individual and ensemble classifiers. Supplemental Figure 3 illustrates the areas under the receiver operating characteristic curves showing prediction performance regarding cluster assignment for the test dataset. In all cases, AUC values were in the good or excellent range.
Discussion
Older individuals derive a survival benefit post-transplant,37 but donor and recipient characteristics may influence the outcomes of both graft failure and mortality after transplantation. In a national cohort, we generated five phenotypes of donors using unsupervised clustering. Recipients of kidneys from two clusters experienced a high risk of all-cause graft failure. While there were differences in the proportion of high-risk donors (defined using the KDRI), donors with high-risk KDRI scores were still found in low-risk clusters.
There are a number of possibilities explaining the tendency of clusters to associate with different recipient outcomes in this study. Clusters 1 and 5 comprised predominantly young, male donors with few comorbidities and the lowest KDRI scores. Recipients of kidneys from both clusters had an intermediate risk of death/graft failure; cluster 1 recipients were at higher risk, which we hypothesize may relate to an increased proportion of Black and obese donors.38,39 Interestingly, although cluster 1 kidneys had some lower risk features (younger donor age, lower risk of hypertension, lower risk of cerebrovascular COD) compared with cluster 3, recipients of these kidneys were still at higher risk of both death with graft function and death-censored graft failure, even after accounting for recipient characteristics. Cluster 4 had the highest proportions of donors with established risk factors for graft failure including obesity,38 hypertension,40 and diabetes.41 Cluster 2 had comparatively fewer donors with high-risk factors, but recipients of kidneys from this cluster were at a comparably high risk of graft loss (including death with graft function and death-censored graft failure). One striking difference is that cluster 2 had the lowest proportion of recipients at an ideal BMI (20%) while cluster 3 had the highest (45%). High donor BMI (>25) is associated with increased surgical retrieval and potentially prolonged warm ischemia time,42,43 which in turn is associated with all-cause graft failure.44 Another possibility is that unique combinations of donor characteristics that drove cluster separation also synergistically conferred risk. Only 5% of donors in cluster 3 had a cerebrovascular COD and elevated BMI (≥30 kg/m2) as opposed to 53% in cluster 2, emphasizing the importance of the inter-relation between factors. For example, BMI may not always represent fat mass, and in the absence of obesity, donors with larger BMIs (relative to recipients) may confer a lower risk of graft loss because of the contribution of more nephrons to the recipient.45 This may explain why those in cluster 3 were at low risk; although 54% of donors in this cluster had an elevated BMI, this may not have reflected obesity. Irrespective of the underlying reasons, our study does suggest that donor characteristics do influence both death and graft failure in older recipients, which is in contrast to some studies that have shown comparable outcomes for older recipients receiving low-risk or medium-risk donor kidneys.13
Recipient characteristics were generally similar across clusters, which may reflect that this was a relatively highly selected population (i.e., a smaller proportion of older adults with kidney failure are candidates for deceased donor transplantation relative to younger patients), further emphasizing the importance of identifying high-risk donor phenotypes that affect recipient outcomes. However, while the identification of differences in outcomes for donor clusters may suggest that future efforts to create prognostic models inclusive of cluster assignment are valuable, this should not be performed without external validation and should not lead to inequitable distribution of organs for older adults.46,47
Recent changes to kidney allocation in the United States have focused on incorporating utility goals such as increasing post-transplant survival rates. Inherent to this goal of utility is to adequately capture donor kidney longevity, ascertained by the KDPI (derived from the KDRI) and inclusive of an opt-in system for older candidates to receive lower-quality donor kidneys.18 Given the association between donor clusters and outcomes for recipients, using cluster assignment may be a novel way to capture organ quality and a means of assigning donor kidneys to older recipients to fit within the updated kidney allocation framework.18 In this regard, a future goal will be to determine whether cluster assignment can be mapped to more contemporary donors across the United States for older recipients. We do acknowledge, however, that a broadly applicable prediction model with good performance characteristics (considering the limited supply of donor organs) would be preferred as opposed to focusing on a particular subgroup. Furthermore, replacing an existing system that broadly attempts to address organ quality for all-comers is not the optimal approach. Therefore, at present, identifying donor phenotypes for older recipients while novel is exploratory and should serve to complement existing measures as opposed to replacing them.
This study has limitations. Our analysis was performed on national registry data; therefore, details on other important outcomes (i.e., changes in kidney function, development of donor-specific antibodies or post-transplant proteinuria) were lacking. Furthermore, we did not have details on other determinants of outcomes for older recipients including immunosuppressive medication use, episodes of acute rejection after transplantation, and patient clinical factors such as frailty.48 Although we addressed missingness with multiple imputation, there is the possibility that our outcomes may have differed with a more complete dataset. Finally, we were pragmatic in our selection of donor variables in cluster derivation and added information on donor organs (including pretransplant biopsy results49) may have better predicted death/graft failure. However, these data were not available in our derivation cohort, and our intended purpose was to use factors readily available in other registry databases.
We used an ML clustering approach inclusive only of donor characteristics and identified clusters of differing donor characteristics for a cohort of older transplant recipients. This approach may be a novel means to characterize donor phenotypes for older adults, acknowledging that donor clusters are associated with differing risks of graft loss in this population. Whether unsupervised clustering can be used to characterize donors to better inform organ allocation for older kidney transplant recipients is an important consideration for future study.
Disclosures
A. Naqvi reports the following—employer: Melville Lodge, GEM Health Care Group. K.K. Tennankore reports the following—consultancy: AstraZeneca, Bayer, GSK, Otsuka, Vifor; research funding: Otsuka Canada; honoraria: AstraZeneca, Baxter, GSK, Otsuka; advisory or leadership role: Associate Editor for the Canadian Journal of Kidney Health and Disease; and speakers bureau: Baxter, Bayer, AstraZeneca. A. Vinson reports the following—consultancy: Paladin Labs Inc. and research funding: Paladin Labs Inc. K. West reports the following—consultancy: Envarsus Canada and honoraria: Envarsus Canada. All remaining authors have nothing to disclose.
Funding
This study was supported through infrastructure funding provided to Dr. Tennankore's research program in his role as the QEII Foundation Endowed Chair in Transplantation Research.
Author Contributions
Conceptualization: Samina Abidi, Syed Sibte Raza Abidi, Karthik K. Tennankore, Kenneth West.
Data curation: Asil Naqvi, Syed Sibte Raza Abidi, Amanda Vinson, George Worthen.
Formal analysis: Asil Naqvi, Syed Sibte Raza Abidi, Karthik K. Tennankore, Amanda Vinson.
Investigation: Bryce Kiberd, Syed Sibte Raza Abidi, Thomas Skinner, Karthik K. Tennankore, Amanda Vinson.
Methodology: Samina Abidi, Bryce Kiberd, Asil Naqvi, Syed Sibte Raza Abidi, Thomas Skinner, Karthik K. Tennankore, Amanda Vinson, George Worthen, Kenneth West.
Project administration: Syed Sibte Raza Abidi, Karthik K. Tennankore.
Resources: Samina Abidi, Karthik K. Tennankore.
Software: Karthik K. Tennankore.
Supervision: Syed Sibte Raza Abidi, Karthik K. Tennankore.
Validation: Asil Naqvi, Syed Sibte Raza Abidi, Karthik K. Tennankore, Amanda Vinson.
Visualization: Asil Naqvi, Syed Sibte Raza Abidi, Karthik K. Tennankore, George Worthen.
Writing – original draft: Asil Naqvi, Syed Sibte Raza Abidi, Karthik K. Tennankore, George Worthen.
Writing – review & editing: Samina Abidi, Bryce Kiberd, Syed Sibte Raza Abidi, Thomas Skinner, Karthik K. Tennankore, Amanda Vinson, Kenneth West.
Supplementary Material
Supplemental Material
This article contains the following supplemental material online at http://links.lww.com/KN9/A371.
Supplemental File 1. Detailed summary of our unsupervised clustering method.
Supplemental Table 1. Baseline characteristics of the study cohort.
Supplemental Figure 1. Proportion of patients experiencing the primary and each secondary outcome.
Supplemental Table 2. Crude outcomes across each donor cluster.
Supplemental Table 3. Time to all-cause graft failure (adjusted model) 1 and 3 years after transplantation. Adjusted for cold ischemia time, recipient factors (body mass index, cause of kidney failure, race, age, functional status, dialysis vintage, height, weight, peak panel reactive antibody status, malignancy, coronary artery disease, peripheral vascular disease, cardiovascular disease, and diabetes), and number of human leukocyte antigen mismatches (PDF).
Supplemental Table 4. Classification performance of individual and ensemble classifier to determine the cluster label (using the test data from the clusters). ROC, receiver operator characteristic score.
Supplemental Figure 2. Area under the curve for the ensemble classifier in the test dataset (overall and for each cluster label).
References
- 1.Jassal SV Krahn MD Naglie G, et al. Kidney transplantation in the elderly: a decision analysis. J Am Soc Nephrol. 2003;14(1):187–196. doi: 10.1097/01.ASN.0000042166.70351.57 [DOI] [PubMed] [Google Scholar]
- 2.Dempster NJ, Ceresa CD, Aitken E, Kingsmore D. Outcomes following renal transplantation in older people: a retrospective cohort study. BMC Geriatr. 2013;13(1):79. doi: 10.1186/1471-2318-13-79 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Jankowska M Bzoma B Malyszko J, et al. Early outcomes and long-term survival after kidney transplantation in elderly versus younger recipients from the same donor in a matched-pairs analysis. Medicine (Baltimore). 2021;100(51):e28159. doi: 10.1097/md.0000000000028159 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Knoll GA. Kidney transplantation in the older adult. Am J Kidney Dis. 2013;61(5):790–797. doi: 10.1053/j.ajkd.2012.08.049 [DOI] [PubMed] [Google Scholar]
- 5.Rao PS, Merion RM, Ashby VB, Port FK, Wolfe RA, Kayler LK. Renal transplantation in elderly patients older than 70 years of age: results from the Scientific Registry of Transplant Recipients. Transplantation. 2007;83(8):1069–1074. doi: 10.1097/01.tp.0000259621.56861.31 [DOI] [PubMed] [Google Scholar]
- 6.Lemoine M Titeca Beauport D Lobbedez T, et al. Risk factors for early graft failure and death after kidney transplantation in recipients older than 70 years. Kidney Int Rep. 2019;4(5):656–666. doi: 10.1016/j.ekir.2019.01.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Pippias M Stel VS Arnol M, et al. Temporal trends in the quality of deceased donor kidneys and kidney transplant outcomes in Europe: an analysis by the ERA-EDTA Registry. Nephrol Dial Transplant. 2021;37(1):175–186. doi: 10.1093/ndt/gfab156 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Rao PS Schaubel DE Guidinger MK, et al. A comprehensive risk quantification score for deceased donor kidneys: the kidney donor risk index. Transplantation. 2009;88(2):231–236. doi: 10.1097/tp.0b013e3181ac620b [DOI] [PubMed] [Google Scholar]
- 9.Clayton PA Dansie K Sypek MP, et al. External validation of the US and UK kidney donor risk indices for deceased donor kidney transplant survival in the Australian and New Zealand population. Nephrol Dial Transplant. 2019;34(12):2127–2131. doi: 10.1093/ndt/gfz090 [DOI] [PubMed] [Google Scholar]
- 10.Rose C, Sun Y, Ferre E, Gill J, Landsberg D, Gill J. An examination of the application of the kidney donor risk index in British Columbia. Can J Kidney Health Dis. 2018;5:2054358118761052. doi: 10.1177/2054358118761052 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Young A Knoll GA McArthur E, et al. Is the kidney donor risk index a useful tool in non-US patients? Can J Kidney Health Dis. 2018;5:2054358118791148. doi: 10.1177/2054358118791148 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Heaphy ELG, Goldfarb DA, Poggio ED, Buccini LD, Flechner SM, Schold JD. The impact of deceased donor kidney risk significantly varies by recipient characteristics. Am J Transplant. 2013;13(4):1001–1011. doi: 10.1111/ajt.12154 [DOI] [PubMed] [Google Scholar]
- 13.Hernandez RA, Malek SK, Milford EL, Finlayson SR, Tullius SG. The combined risk of donor quality and recipient age: higher-quality kidneys may not always improve patient and graft survival. Transplantation. 2014;98(10):1069–1076. doi: 10.1097/tp.0000000000000181 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Shang N Khan A Polubriaginof F, et al. Medical records-based chronic kidney disease phenotype for clinical care and “big data” observational and genetic studies. NPJ Digit Med. 2021;4(1):70. doi: 10.1038/s41746-021-00428-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Su C Zhang Y Flory JH, et al. Clinical subphenotypes in COVID-19: derivation, validation, prediction, temporal patterns, and interaction with social determinants of health. NPJ Digit Med. 2021;4(1):110. doi: 10.1038/s41746-021-00481-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Zhang XA Yates A Vasilevsky N, et al. Semantic integration of clinical laboratory tests from electronic health records for deep phenotyping and biomarker discovery. NPJ Digit Med. 2019;2(1):32. doi: 10.1038/s41746-019-0110-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Allen C, Hu J, Kumar V, Ahmad M, Teredesai A. Interpretable Phenotyping for Electronic Health Records. 9th IEEE International Converence on Healthcare Informatics. Canada; 2021:161–170. [Google Scholar]
- 18.Friedewald JJ Samana CJ Kasiske BL, et al. The kidney allocation system. Surg Clin North Am. 2013;93(6):1395–1406. doi: 10.1016/j.suc.2013.08.007 [DOI] [PubMed] [Google Scholar]
- 19.Stegall MD, Stock PG, Andreoni K, Friedewald JJ, Leichtman AB. Why do we have the kidney allocation system we have today? A history of the 2014 kidney allocation system. Hum Immunol. 2017;78(1):4–8. doi: 10.1016/j.humimm.2016.08.008 [DOI] [PubMed] [Google Scholar]
- 20.Stewart DE Wilk AR Toll AE, et al. Measuring and monitoring equity in access to deceased donor kidney transplantation. Am J Transplant. 2018;18(8):1924–1935. doi: 10.1111/ajt.14922 [DOI] [PubMed] [Google Scholar]
- 21.Bzdok D, Altman N, Krzywinski M. Statistics versus machine learning. Nat Methods. 2018;15(4):233–234. doi: 10.1038/nmeth.4642 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Rajula HSR, Verlato G, Manchia M, Antonucci N, Fanos V. Comparison of conventional statistical methods with machine learning in medicine: diagnosis, drug development, and treatment. Medicina (Kaunas). 2020;56(9):455. doi: 10.3390/medicina56090455 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Thongprayoon C Vaitla P Jadlowiec CC, et al. Use of machine learning consensus clustering to identify distinct subtypes of Black kidney transplant recipients and associated outcomes. JAMA Surg. 2022;157(7):e221286. doi: 10.1001/jamasurg.2022.1286 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Mao SA Thongprayoon C Jadlowiec CC, et al. Distinct Phenotypes of Kidney Transplant Recipients Aged 80 Years or Older in the United States by Machine Learning Consensus Clustering. 2022 American Transplant Congress. 2022. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Miller AJ, Kiberd BA, Alwayn IP, Odutayo A, Tennankore KK. Donor-recipient weight and sex mismatch and the risk of graft loss in renal transplantation. Clin J Am Soc Nephrol. 2017;12(4):669–676. doi: 10.2215/Cjn.07660716 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Swanson EA Patel MS Groat T, et al. Vasopressor selection during critical care management of brain dead organ donors and the effects on kidney graft function. J Trauma Acute Care Surg. 2020;88(6):783–788. doi: 10.1097/ta.0000000000002688 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Leeaphorn N, Garg N, Thamcharoen N, Khankin EV, Cardarelli F, Pavlakis M. Cytomegalovirus mismatch still negatively affects patient and graft survival in the era of routine prophylactic and preemptive therapy: a paired kidney analysis. Am J Transplant. 2019;19(2):573–584. doi: 10.1111/ajt.15183 [DOI] [PubMed] [Google Scholar]
- 28.Wang W, Arora R, Livescu K, Bilmes J. On deep multi-view representation learning. Proceedings of the 32nd International Conference on Machine Learning. PMLR; 2015:1083–1092. [Google Scholar]
- 29.Mitra S, Saha S, Hasanuzzaman M. Multi-view clustering for multi-omics data using unified embedding. Sci Rep. 2020;10(1):13654. doi: 10.1038/s41598-020-70229-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Ji J, Li R, Pang W, He F, Feng G, Zhao X. A multi-view clustering algorithm for mixed numeric and categorical data. IEEE Access. 2021;9:24913–24924. doi: 10.1109/access.2021.3057113 [DOI] [Google Scholar]
- 31.Bickel SST. Multi-view clustering. Proceedings—Fourth IEEE International Conference on Data Mining. Brighton; 2004:19–26. [Google Scholar]
- 32.Lim WH, Chapman JR, Wong G. Peak panel reactive antibody, cancer, graft, and patient outcomes in kidney transplant recipients. Transplantation. 2015;99(5):1043–1050. doi: 10.1097/tp.0000000000000469 [DOI] [PubMed] [Google Scholar]
- 33.Rendón E, Abundez I, Arizmendi A, Quiroz E. Internal versus external cluster validation indexes. Int J Comput Commun. 2011;5:27–34. [Google Scholar]
- 34.Agrawal U Soria D Wagner C, et al. Combining clustering and classification ensembles: a novel pipeline to identify breast cancer profiles. Artif Intell Med. 2019;97:27–37. doi: 10.1016/j.artmed.2019.05.002 [DOI] [PubMed] [Google Scholar]
- 35.Purandhar N, Ayyasamy S, Siva Kumar P. Classification of clustered health care data analysis using generative adversarial networks (GAN). Soft Comput. 2022;26(12):5511–5521. doi: 10.1007/s00500-022-07026-7 [DOI] [Google Scholar]
- 36.Ambler G, Omar RZ, Royston P. A comparison of imputation techniques for handling missing predictor values in a risk model with a binary outcome. Stat Methods Med Res. 2007;16(3):277–298. doi: 10.1177/0962280206074466 [DOI] [PubMed] [Google Scholar]
- 37.Wong G Howard K Chapman JR, et al. Comparative survival and economic benefits of deceased donor kidney transplantation and dialysis in people with varying ages and co-morbidities. PLoS One. 2012;7(1):e29591. doi: 10.1371/journal.pone.0029591 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Naik AS Zhong Y Parasuraman R, et al. The temporal and long-term impact of donor body mass index on recipient outcomes after kidney transplantation—a retrospective study. Transplant Int. 2020;33(1):59–67. doi: 10.1111/tri.13505 [DOI] [PubMed] [Google Scholar]
- 39.Molnar MZ Kovesdy CP Bunnapradist S, et al. Donor race and outcomes in kidney transplant recipients. Clin Transplant. 2013;27(1):37–51. doi: 10.1111/j.1399-0012.2012.01686.x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Altheaby A Al Dalbhi S Alghamdi Y, et al. Effect of donor hypertension on renal transplant recipients' blood pressure, allograft outcomes and survival: a systematic review and meta-analysis. Am J Cardiovasc Dis. 2019;9(4):49–58. [PMC free article] [PubMed] [Google Scholar]
- 41.Cohen JB, Bloom RD, Reese PP, Porrett PM, Forde KA, Sawinski DL. National outcomes of kidney transplantation from deceased diabetic donors. Kidney Int. 2016;89(3):636–647. doi: 10.1038/ki.2015.325 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Arshad A Hodson J Chappelow I, et al. The impact of donor body mass index on outcomes after deceased kidney transplantation—a national population-cohort study. Transplant Int. 2018;31(10):1099–1109. doi: 10.1111/tri.13263 [DOI] [PubMed] [Google Scholar]
- 43.Vinson AJ Rose C Kiberd BA, et al. Factors associated with prolonged warm ischemia time among deceased donor kidney transplant recipients. Transplant Direct. 2018;4(5):e342. doi: 10.1097/txd.0000000000000781 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Tennankore KK, Kim SJ, Alwayn IP, Kiberd BA. Prolonged warm ischemia time is associated with graft failure and mortality after kidney transplantation. Kidney Int. 2016;89(3):648–658. doi: 10.1016/j.kint.2015.09.002 [DOI] [PubMed] [Google Scholar]
- 45.McGee J Magnus JH Islam TM, et al. : Donor-recipient gender and size mismatch affects graft success after kidney transplantation. J Am Coll Surg. 2010;210(5):718–725.e1. doi: 10.1016/j.jamcollsurg.2009.12.032 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Clement J, Maldonado AQ. Augmenting the transplant team with artificial intelligence: toward meaningful AI use in solid organ transplant. Front Immunol. 2021;12:694222. doi: 10.3389/fimmu.2021.694222 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Babic BI, Cohen G, Evgeniou T, Gerke S, Trichakis N. Can AI fairly decide who gets an organ transplant? Harvard Business Review. 2020. [Google Scholar]
- 48.Worthen G, Tennankore K. Frailty screening in chronic kidney disease: current perspectives. Int J Nephrol Renovasc Dis. 2019;12:229–239. doi: 10.2147/ijnrd.s228956 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 49.Luo Y Liang J Hu X, et al. Deep learning algorithms for the prediction of posttransplant renal function in deceased-donor kidney recipients: a preliminary study based on pretransplant biopsy. Front Med (Lausanne). 2021;8:676461. doi: 10.3389/fmed.2021.676461 [DOI] [PMC free article] [PubMed] [Google Scholar]