Discussion
Previous studies analysing the UNOS dataset of kidney transplant recipients ≥80 years old described the overall patient characteristics as primarily white males (>80%) who received a kidney (from donors approximately 50 years in age) with a mean cold ischaemia time of 16.72 hours.24 25 By applying an ML approach, our present study demonstrates greater heterogeneity among this patient population, and our ML consensus cluster analysis successfully identified three clusters of very elderly kidney transplant recipient with unique clinical characteristics and associated post-transplant outcomes.
Cluster 1 recipients were more likely to be on dialysis at the time of transplant, and receive a deceased donor kidney transplant from a younger non-ECD KDPI <85% male donor without hypertension. Recipients in cluster 2 were also more likely to be on dialysis at the time of transplant. Cluster 2 recipients however received deceased donor kidneys from older, hypertensive, ECD donors with KDPI scores ≥85%. Cluster 2 recipients had a greater number of HLA mismatches, the longest cold ischaemia time and the highest use of machine perfusion for the transplanted kidney. In contrast, cluster 3 recipients were more likely to either be preemptive or be on dialysis for less than 1 year prior to kidney transplant. Cluster 3 recipients received living donor kidney transplants, had the lowest number of HLA mismatches and were more likely to receive nondepleting induction therapy. Glomerular disease was the most common indication for transplant in all clusters (43%). Recipients in all clusters also had good functional status with an overall low incidence of diabetes and peripheral vascular disease. Acute rejection events were low in all clusters. Among the three clusters, cluster 3 had the most favourable post-transplant outcomes specific to patient survival and death-censored graft failure.
Patients in all clusters were unlikely to be sensitised (low PRA). The number of HLA mismatches was highest in clusters 1 and 2 and lowest in cluster 3, which is likely a reflection of living-related kidney donation. Despite the low PRA in clusters 1 and 2, depleting induction (thymoglobulin, alemtuzumab) was used in a significant number of recipients (table 1). Overall rates of acute rejection were low in all clusters ranging from 0% in cluster 3 to 6.6% in cluster 2. Cluster 2 patients had the highest acute rejection rates at 1 year. It can be hypothesised that this may be due to the higher number of HLA mismatches, longer cold ischaemia time and higher occurrence of DGF compared with the other clusters. Cluster 3 recipients had the lowest incidence of acute rejection (0%) at 1 year among all clusters despite more patients receiving non-depleting induction (basiliximab) and steroid-sparing regimens. These recipients had the lowest number of HLA mismatches and DGF, and these factors may have contributed to the low incidence of rejection. It has been suggested that the risk for acute rejection is lower in older kidney transplant recipients due to immunosenescence. Recent evidence suggests that lower-intensity immunosuppression regimens (steroid-sparing) offer beneficial outcomes in older kidney transplant recipients by balancing risk for rejection and also helping to minimise immunosuppression side effects.45 Among very elderly kidney transplant recipients aged ≥80, the findings of our ML consensus cluster analysis confirms that lower-intensity immunosuppression regimens appears safe with low risk of rejection for patients with clinical characteristics shown in clusters 1 and 3.
Key features for cluster 3 recipients included preemptive kidney transplantation or dialysis duration for less than 1 year prior kidney transplant. The majority of recipients in cluster 3 received living donor kidney transplants from non-hypertensive donors. Transplantation in the elderly, particularly for those beyond the age of 80 years, is often a strongly debated topic that takes into account benefit to the recipient and judicious use of a scare resource.25–28 For elderly recipients, there can be additional controversy over the use of a living donor kidney due to overall lower survival as a general limitation of age and comorbidities.28 In this study, cluster 3 recipients had the best survival, 96.9% at 1 year, and the lowest risk for graft loss. Hypertension and diabetes, generally account for a significant proportion of kidney transplants in the elderly.16 46 In our analysis, glomerular disease was the most common kidney disease aetiology (43%) for all three clusters. Moreover, comorbidities such as diabetes (28%) and peripheral vascular disease (11%) were less common and the majority of recipients had a Karnofsky functional score ranging from 80%–100%. These data suggest that the recipient profile for kidney transplant recipients greater than 80 years of age differs to that of all kidney transplant recipients aged greater than 65 years of age. Overall, 1-year and 5-year patient survival was 93% and 66% in cluster 1, 88% and 46% in cluster 2, and 97% and 61% in cluster 3. Recipients in clusters 1 and 3 had comparable patient survival, and cluster 2 patients had the highest mortality, despite comparable factors between cluster 1 and 2 for post-transplant patient survival in older kidney transplant recipients such as recipients age, dialysis vintage, comorbidities and functional status.47 48
One-year and five year death-censored graft survival was 96% and 91% in cluster 1, 90% and 81% in cluster 2, and 100% and 98% in cluster 3, respectively. Cluster 2 patients had the highest death-censored graft failure at 5 years among the clusters. The highest of proportion of cluster 2 recipients were in UNOS Region 1. Lower graft survival in cluster 2 may be explained by donor quality as these recipients were more likely to receive deceased donor kidney transplants from hypertensive ECD deceased donors with a KDPI score ≥85%. Kidneys from donor with a KDPI score greater than >85%, so-called high KDPI kidneys, are known to have shorter graft survival, and are often reserved for older recipients who have comorbidities, like diabetes and cardiovascular disease, that otherwise limit their long-term survival.16 In this study, the overall incidence of diabetes and peripheral vascular disease was lower for recipients above the age of 80 years and the primary indication for transplant was glomerular disease. As such, patients above the age of 80 years who meet criteria to qualify for a transplant may be less likely to have other common comorbidities, such as diabetes, and be self-selected to have better survival that extends beyond the standard survival for an elderly kidney transplant recipient and a high KDPI kidney. In our ML clustering analysis, both cluster 3 recipients who were more likely to received preemptive living donor kidney transplants and cluster 1 recipients receiving standard non-ECD deceased donor kidney transplants had better death-censored graft survival than cluster 2 recipients.
Our study has several limitations. The UNOS database has inherent limitations including lack of granular data regarding cause of patient death and graft loss. In addition, we applied ML cluster analysis on a retrospectively reported multicentre database. All very elderly kidney transplant recipients have undergone a comprehensive pretransplant evaluation, however, each transplant programme has differing criteria for the management of patients prior, during and after transplant.49 Furthermore, some transplant programmes currently offer kidney transplantation only to older candidates with living donors due to concern of waitlist mortaility and perioperative morbidity and mortality.50 51 Third, while it is possible that missing data were not completely random, all variables in our study had missing data <5%. Therefore, it is unlikely that missing data imputation would substantially alter the result of our analysis. In addition, unlike supervised ML that data model bias is a challenge, an unsupervised learning clustering algorithm has parameters that control the model’s flexibility to fit the data and can learn bias from dataset. Nevertheless, unsupervised models can still encounter particular biases in data composition. Thus, the potential gender and racial bias based on populations’ ethnic backgrounds or geographical locations should be noted. Lastly, data on quality-of-life post-transplant are limited in the UNOS database.52 53 While cluster 2 had the worst post-transplant outcomes among all clusters, future studies assessing the quality of life in this cluster of very elderly kidney transplant recipients aged ≥80 are needed.
Despite limitations, our study using an unsupervised ML clustering approach identifies distinct clusters within kidney transplant recipients aged ≥80 years. While the survival benefits of kidney transplant in octogenarians have previously been compared with remaining on dialysis,7 12 22 24 54 the findings of our study provide further insights into the different allograft and patient outcomes among the unique phenotypic subtypes of very elderly kidney transplant recipients. Glomerular disease was the most common kidney disease aetiology (43%) for all clusters and comorbidities such as diabetes (28%) and peripheral vascular disease (11%) were less common. The majority of recipients in all clusters had a Karnofsky functional score ranging from 80% to 100%. Despite advanced age, cluster 3 recipients, who were largely preemptive and received living donor kidney transplants, had favourable outcomes (both allograft and patient survival) comparable to younger kidney transplant recipients. Compared with cluster 3, cluster 1 recipients had comparable survival but higher death-censored graft failure, while recipients in cluster 2 had the worst post-transplant outcomes specific to patient survival and death-censored graft failure. Cluster 2 recipients also had the highest incidence of acute rejection (6.6% vs 2.0% and 0%). Future studies are required to better identify specific differences between cluster 2 recipients, who had less than 50% survival at 5 years post-transplant, compared with recipients in clusters 1 and 3 so as to better guide clinical and patient decision making specific to transplant. In addition, while the findings of unsupervised ML clustering approach in this study provide detailed information on distinct phenotypes of kidney recipients aged ≥80 in the USA and associated outcomes with differing post-transplant outcomes, ML clustering algorithms have their limitations that do not directly generate risk prediction for each individual. Thus, future studies assessing the utilisation of supervised ML prediction models for transplant outcomes among kidney transplant recipients ≥80 in the USA are required.
The ML clustering approach produced three phenotypic clusters of very elderly kidney transplant recipients aged ≥80 in the USA. Post-transplant outcomes differed among the clusters including variability in allograft rejection, allograft loss and patient mortality. Our study also demonstrated a varying geographical distribution of kidney recipients aged ≥80 in the USA in the different UNOS Regions in the USA. Our approach identifies targets for individualised medicine and opportunities to improve care for very elderly kidney transplant recipients, particularly those within cluster 2 subtype.
