Visual Abstract
Keywords: dialysis
Abstract
Key Points
In a contemporary UK cohort of 96,809 patients, peritoneal dialysis was associated with lower mortality than hemodialysis across all phenotypes.
The magnitude of this association varied by phenotype and attenuated with increasing age.
Phenotype-based analyses may support more individualized dialysis modality selection and inform system-level policy decisions.
Background
Whether peritoneal dialysis (PD) is associated with lower mortality than hemodialysis in contemporary practice remains debated, particularly across heterogeneous patient populations and in the presence of competing transplantation. We examined modality-associated mortality differences across data-driven phenotypes of incident dialysis patients in the United Kingdom using a competing-risks framework.
Methods
We analyzed 96,809 adults initiating dialysis in the United Kingdom Renal Registry between 2007 and 2021. Unsupervised k-prototypes clustering was used to derive phenotypes based on age, sex, ethnicity, primary kidney disease, hemoglobin, serum albumin, and transplant-listing status. Mortality during the dialysis phase before transplantation was analyzed using competing-risks methods, treating kidney transplantation as a competing event. Cumulative incidence of death at 1 and 5 years was modeled using jack-knife pseudo-value regression with complementary log-log links, adjusting for demographic and clinical covariates. Dialysis modality (hemodialysis versus PD) was the primary exposure, with stratified analyses performed within each phenotype. A complementary competing-risks analysis examined time to transplantation.
Results
Among incident patients, 24% initiated PD and 76% hemodialysis. Three reproducible phenotypes were identified, differing primarily by age, hemoglobin and albumin levels, and transplant-listing status. Across the overall cohort, hemodialysis was associated with a higher cumulative incidence of death before transplantation at both 1 year (subdistribution hazard ratio [sHR], 1.85; 95% confidence interval [CI], 1.75 to 1.96) and 5 years (sHR, 1.47; 95% CI, 1.43 to 1.51). In stratified analyses, hemodialysis was associated with higher mortality across all phenotypes. At 1 year, adjusted sHRs ranged from 1.55 (95% CI, 1.44 to 1.68) to 2.29 (95% CI, 1.87 to 2.80) across clusters. At 5 years, the association persisted but attenuated with increasing age, with adjusted sHRs of 2.42 (95% CI, 2.23 to 2.63) in the youngest phenotype, 1.55 (95% CI, 1.48 to 1.61) in the intermediate phenotype, and 1.14 (95% CI, 1.11 to 1.19) in the oldest phenotype. In complementary analyses, time to transplantation did not differ significantly across phenotypes after adjustment.
Conclusions
In contemporary UK practice, hemodialysis was associated with higher mortality during the dialysis phase compared with PD across distinct patient phenotypes, with time-dependent variation in effect magnitude. These registry-based associations support consideration of phenotype-informed modality selection while acknowledging the potential for residual confounding and the influence of health system context.
Introduction
Whether and under which circumstances peritoneal dialysis (PD) offers a survival advantage over hemodialysis remains one of the longest-running debates in nephrology. Earlier registry studies from the 1990s and early 2000s—conducted before widespread use of biocompatible PD fluids, automated cyclers, and high-flux hemodialysis membranes—reported conflicting results, with any early PD benefit often diminishing after 1–2 years, particularly among older adults.1,2 These findings contributed to a persistent perception that PD is less suitable for elderly or comorbid patients, despite substantial advances in dialysis technology and patient management over the past two decades.
A recent Cochrane review reported a pooled mortality risk ratio of 0.87 (95% confidence interval [CI], 0.77 to 0.98) favoring PD, but with extreme heterogeneity (I2=99%) and a very low certainty due to reliance on older observational cohorts.3 Even UK registry analyses predate many of today's improvements, leaving uncertainty about whether earlier patterns persist in contemporary practice and obscuring potential differences across patient subgroups.
Machine-learning methods such as k-prototypes clustering offer a data-driven way to define reproducible phenotypes of incident dialysis patients using mixed demographic, clinical, and laboratory features. By stratifying patients into phenotypes, survival can be compared between PD and hemodialysis within clinically homogeneous groups, moving beyond one-size-fits-all estimates toward personalized modality selection.4
At the same time, modality uptake is shaped not only by patient factors but also by health system context. In the United Kingdom, only around one in four patients start PD, raising concerns about an evidence-practice gap. Differences in incidence, prevalence, infrastructure, and costs between the United Kingdom and Western Europe may further influence modality use.
In this study, we analyzed a contemporary UK Renal Registry (UKRR) cohort using unsupervised clustering to compare PD and hemodialysis survival across phenotypes. We applied an unsupervised (data-driven) clustering approach to identify naturally occurring phenotypic subgroups among incident dialysis patients. In this context, “unsupervised” indicates that the grouping was based solely on baseline clinical characteristics and did not incorporate survival outcomes. To contextualize these findings, we drew on the International Society of Nephrology (ISN) Global Kidney Health Atlas (GKHA) to compare the United Kingdom with Western Europe for disease burden, dialysis infrastructure, and treatment costs.5–7 Our aim was to generate evidence that is both clinically informative and policy-relevant, bridging patient-level outcomes with system-level realities.
Methodology
Data Source and Study Cohort
We obtained data on all adults (≥18 years) initiating dialysis in the United Kingdom from January 1, 2007, through September 30, 2021, via the UKRR. The UKRR aggregates quarterly submissions from every kidney unit in England, Wales, and Northern Ireland, capturing patient demographics, primary kidney disease, dialysis modality changes, transplant listing status, and key laboratory values (e.g., hemoglobin, albumin). Hemoglobin and serum albumin were analyzed as baseline variables, defined as the values recorded in the UKRR quarterly submission immediately preceding dialysis initiation, and were not modeled as time-varying covariates. Local units validate any flagged inconsistencies before upload, yielding near-complete recording of diagnosis (>99%) and over 60% completeness for other clinical fields. After excluding entries with missing core data, our analysis cohort comprised 96,809 deidentified patients, under a data-sharing agreement and ethical approval from the UKRR (DSA125).
Each patient contributed follow-up from the date they began either PD or hemodialysis until the earliest of: a sustained (>0 days) switch in dialysis modality, kidney transplantation, loss to follow-up, or death. We defined the technique-failure date as the first occasion on which a patient remained on an alternate modality for more than 30 days. Baseline variables included age at dialysis initiation, sex, White versus non-White ethnicity, primary kidney disease etiology (polycystic, pyelonephritis, GN, renovascular, other known causes, or unknown), transplant-listing status, hemoglobin (g/L), and serum albumin (g/L). Variables with more than 20% missing data—such as other comorbidities and additional laboratory measurements—were omitted to avoid reliance on extensive imputation and preserve statistical power. Dialysis modality was defined at initiation and analyzed on an intention-to-treat basis; patients were not reclassified according to subsequent modality changes, and follow-up was administratively censored at sustained modality switch rather than reassigned to the new modality.
Phenotype Derivation via k-Prototypes Clustering
We used the k-prototypes algorithm to partition patients into homogenous “phenotypes” based on the seven baseline variables available in the registry, with missing data <20% (age, sex, ethnicity, primary kidney disease, hemoglobin, albumin, transplant-listing).
K-prototype clustering is an unsupervised machine-learning method designed to group individuals using a combination of continuous and categorical variables. In this study, it allowed us to classify patients into clinically meaningful phenotypes based on age, laboratory values, primary kidney disease, demographic characteristics, and transplant-listing status. Patients within each cluster therefore share broadly similar clinical profiles. We evaluated solutions with three to six clusters and selected three clusters based on standard measures of cluster separation and interpretability.8 Cluster stability was assessed by repeating the procedure with different random starting points and confirming consistent patient assignment across runs.
Confounding Control and Role of Phenotype Clustering
Dialysis modality selection is influenced by multiple patient, clinician, and health system factors. The k-prototype clustering approach was used to identify clinically coherent phenotypes and to examine heterogeneity in the association between dialysis modality and mortality across patient groups; clustering was not intended to control for confounding. Confounding was addressed through multivariable adjustment within the competing-risks regression models, which included demographic, clinical, and laboratory variables known to influence both dialysis modality selection and mortality, including age, sex, ethnicity, primary kidney disease, hemoglobin, serum albumin, and transplant-listing status. Residual confounding due to unmeasured determinants of modality choice cannot be excluded.
Competing-Risks Modelling
Mortality was evaluated using a competing-risks framework, with kidney transplantation treated as the competing event. The estimand of interest was mortality before transplantation. Cumulative incidence functions for death were estimated using the Fine-Gray subdistribution approach, which appropriately accounts for competing events by retaining individuals who undergo transplantation within the risk set.
To obtain marginal, population-level estimates of mortality at clinically relevant time horizons, we calculated jack-knife pseudo-values for the cumulative incidence of death at 1 and 5 years after dialysis initiation. This method first calculates the cumulative incidence of death at a given time horizon (e.g., 5 years) while accounting for transplantation as a competing event. Individual pseudo-values derived from this cumulative incidence are then modeled using generalized linear models, allowing estimation of population-level, covariate-adjusted subdistribution hazard ratios (sHRs). This approach provides clinically interpretable estimates of mortality risk at specific time points rather than relying solely on instantaneous hazard ratios.
These pseudo-values were modeled using generalized linear models with a complementary log-log link, adjusting for baseline demographic and clinical variables included in the clustering analysis (age, sex, ethnicity, primary kidney disease, hemoglobin, and serum albumin). Transplant-listing status was not included as an adjustment variable in the mortality models, as it lies on the causal pathway to kidney transplantation, which was treated as a competing event; conditioning on listing would therefore change the estimand from population-level mortality before transplantation to mortality conditional on transplant eligibility and access.
To evaluate whether the association between dialysis modality and mortality differed across phenotypes, an interaction term between dialysis modality and phenotype (modality×phenotype) was included in the pseudo-value regression models. This allowed formal testing for effect heterogeneity across the three phenotypic groups. As a sensitivity analysis, models were also fitted separately within each phenotype at each time horizon.
Cumulative incidence functions for transplantation were estimated using the Fine-Gray subdistribution approach. Jack-knife pseudo-values were calculated for the cumulative incidence of transplantation at 1 and 5 years following dialysis initiation and modeled using generalized linear models with a complementary log-log link. Models were adjusted for age, sex, ethnicity, primary kidney disease, hemoglobin, serum albumin, and transplant-listing status, with phenotype membership included to evaluate differences in access to transplantation across phenotypes. Transplant-listing status was included in these models to condition on baseline eligibility for transplantation and to distinguish access to transplantation from competing mortality risk.
This mirror analysis was used to describe phenotype-specific transplantation dynamics and to contextualize differences in mortality before transplantation observed across dialysis modalities and phenotypic groups.
Model Performance
Model discrimination was assessed using the concordance index (C-index) applied to predicted cumulative incidence values. To avoid optimistic bias, C-indices were estimated using cross-validation, with model derivation and evaluation performed on separate data splits.
Cluster Reproducibility
To ensure that our k-prototype solution was not driven by a particular random initialization, we repeated the clustering procedure ten times (k=3, γ=1) using distinct random seeds (0–9). For each pair of runs, we computed the adjusted rand index (ARI) to quantify agreement in cluster membership, resulting in 45 pairwise ARI values.9 We then summarized these by reporting the average and full range of ARIs across all run-run comparisons. High mean ARI values (close to 1) indicated that our clusters were stable and not an artifact of the starting centroids.
Health System Context
To place our findings in international context, we extracted health system data for the United Kingdom and Western Europe from the 2023 ISN GKHA. Variables included the incidence and prevalence of kidney failure (per million population), dialysis facility density (hemodialysis and PD centers per million population), transplant center density, and annual costs of in-center hemodialysis, PD, and kidney transplantation (first year and subsequent maintenance years). For each variable, the UK country-specific values were compared with the Western European regional medians reported by the ISN-GKHA. Temporal trends since 2019 were also extracted where available.7
Statistical Software
Data management and phenotype derivation (k-prototype clustering and cluster stability assessment) were performed in Python (v3.10) using pandas and the kmodes package.10 Competing-risks analyses, including Fine-Gray cumulative incidence estimation, jack-knife pseudo-value generation, and pseudo-value regression using generalized linear models with a complementary log-log link (with robust standard errors and bootstrap resampling), were performed in Stata (v18, StataCorp).
Ethics Approval
This study used pseudonymized data provided by the UKRR under an approved data sharing agreement (DSA125). The study was conducted in accordance with ethical approval granted by the Research Ethics Committee (Ref: 21/NE/0045). The UKRR holds permissions under Section 251 of the National Health Service Act 2006 (Ref: 16/CAG/0064) to collect and share confidential patient data for research purposes, with overarching ethical approval from the Research Ethics Committee (Ref: 16/NE/0042). Individual patient consent was not required.
Results
Baseline Characteristics by Dialysis Modality
Of 96,809 incident dialysis patients, 22,711 (23.5%) initiated PD and 74,098 (76.5%) initiated hemodialysis (Table 1). The proportion of female patients was similar between modalities (37.0% for PD versus 36.5% for hemodialysis), and racial composition was broadly comparable, with most patients identifying as White (72.5% versus 68.7%), followed by Asian (11.6% versus 11.1%) and Black ethnicity (6.0% versus 6.8%); mixed, other, or unknown ethnicity accounted for the remainder.
Table 1.
Baseline characteristics of the peritoneal dialysis and hemodialysis cohorts
| Characteristic | Overall (n=96,809) | PD (n=22,711) | Hemodialysis (n=74,098) |
|---|---|---|---|
| Age, yr, mean (SD) | 63.0 (15.5) | 58.6 (16.4) | 64.5 (15.0) |
| Hemoglobin, g/L, mean (SD) | 100.0 (15.0) | 107.9 (15.2) | 97.4 (14.4) |
| Albumin, g/L, mean (SD) | 33.5 (6.6) | 34.7 (6.3) | 33.1 (6.6) |
| Female sex, n (%) | 35,465 (36.6) | 8397 (37.0) | 27,068 (36.5) |
| Transplant listed, n (%) | 31,936 (33.0) | 12,271 (54.0) | 19,665 (26.5) |
| Ethnicity, n (%) | |||
| Asian | 10,871 (11.2) | 2631 (11.6) | 8240 (11.1) |
| Black | 6402 (6.6) | 1370 (6.0) | 5032 (6.8) |
| Mixed | 997 (1.0) | 263 (1.2) | 734 (1.0) |
| Other | 1308 (1.4) | 299 (1.3) | 1009 (1.4) |
| Unknown | 9908 (10.2) | 1698 (7.5) | 8210 (11.1) |
| White | 67,323 (69.5) | 16,450 (72.5) | 50,873 (68.7) |
| Primary kidney disease, n (%) | |||
| Diabetic kidney disease | 26,025 (26.9) | 5721 (25.2) | 20,304 (27.4) |
| GN | 11,646 (12.0) | 3760 (16.6) | 7886 (10.6) |
| Hypertensive nephropathy | 6511 (6.7) | 1635 (7.2) | 4876 (6.6) |
| Polycystic kidney disease | 5430 (5.6) | 1882 (8.3) | 3548 (4.8) |
| Pyelonephritis | 5553 (5.7) | 1273 (5.6) | 4280 (5.8) |
| Renovascular disease | 6061 (6.3) | 1139 (5.0) | 4922 (6.6) |
| Other | 16,320 (16.9) | 3069 (13.5) | 13,251 (17.9) |
| Unknown/blank | 19,263 (19.9) | 4232 (18.6) | 15,031 (20.3) |
Patients initiating PD were younger than those initiating hemodialysis (mean age 58.6 [16.4] versus 64.5 [15.0] years) and had higher mean hemoglobin (107.9 [15.2] versus 97.4 [14.4] g/L) and serum albumin concentrations (34.7 [6.3] versus 33.1 [6.6] g/L). A greater proportion of patients with PD were listed for kidney transplantation at baseline compared with hemodialysis patients (54% versus 27%).
Regarding primary kidney disease, GN (16.6% versus 10.6%) and polycystic kidney disease (8.3% versus 4.8%) were more common among PD starters, whereas diabetic kidney disease (25.2% versus 27.4%) and renovascular disease (5.0% versus 6.6%) were more frequent among hemodialysis starters. Other or unknown causes accounted for approximately one third of patients in each modality. These findings indicate substantial baseline case-mix differences between PD and hemodialysis cohorts, motivating subsequent phenotype-based analyses.
Comparison between Phenotypic Clusters
Unsupervised k-prototype clustering identified three phenotypes with distinct baseline characteristics (Table 2).
Table 2.
Comparison between clusters 1, 2, and 3 for baseline characteristics
| Variable | Cluster 1 | Cluster 2 | Cluster 3 |
|---|---|---|---|
| Hemoglobin (g/L) | 95.1 (12.9) | 116.1 (9.9) | 91.1 (9.5) |
| Albumin (g/L) | 34.3 (6.7) | 35.3 (6.0) | 31.6 (6.4) |
| Age (yr) | 43.1 (10.1) | 65.6 (10.6) | 73.6 (7.9) |
| Sex | |||
| Female, n (%) | 9738 (38.7) | 10,811 (35.5) | 14,916 (36.2) |
| Male, n (%) | 15,414 (61.3) | 19,658 (64.5) | 26,272 (63.8) |
| Ethnicity | |||
| Asian, n (%) | 3609 (15.7) | 3259 (11.9) | 4003 (11.0) |
| Black, n (%) | 2820 (12.3) | 1528 (5.6) | 2054 (5.6) |
| Mixed, n (%) | 414 (1.8) | 280 (1.0) | 303 (0.8) |
| Other, n (%) | 502 (2.2) | 387 (1.4) | 419 (1.1) |
| White, n (%) | 15,621 (68.0) | 21,937 (80.1) | 29,765 (81.4) |
| Primary kidney disease | |||
| Diabetic kidney disease, n (%) | 7008 (32.7) | 8506 (34.4) | 10,511 (33.5) |
| GN, n (%) | 4461 (20.8) | 3633 (14.7) | 3552 (11.3) |
| Hypertensive nephropathy, n (%) | 1684 (7.9) | 2149 (8.7) | 2678 (8.5) |
| Other causes, n (%) | 4573 (21.4) | 4347 (17.6) | 7400 (23.6) |
| Polycystic kidney disease, n (%) | 1924 (9.0) | 2347 (9.5) | 1159 (3.7) |
| Pyelonephritis, n (%) | 1446 (6.8) | 1669 (6.8) | 2438 (7.8) |
| Renovascular disease, n (%) | 319 (1.5) | 2062 (8.3) | 3680 (11.7) |
| Transplant listing | |||
| Not listed, n (%) | 8683 (34.5) | 19,739 (64.8) | 36,451 (88.5) |
| Listed, n (%) | 16,469 (65.5) | 10,730 (35.2) | 4737 (11.5) |
Cluster 1 comprised the youngest patients (mean age 43.1 [10.1] years), with intermediate hemoglobin and albumin levels (95.1 [12.9] and 34.3 [6.7] g/L). This phenotype had the highest proportions of Asian (15.7%) and Black (12.3%) patients and the lowest proportion of White patients (68.0%). GN (20.8%) and polycystic kidney disease (9.0%) were more common in this cluster, and 65.5% of patients were listed for transplantation at baseline.
Cluster 2 represented an intermediate-age phenotype (65.6 [10.6] years) with the highest hemoglobin and albumin concentrations (116.1 [9.9] and 35.3 [6.0] g/L). Patients were predominantly White (80.1%), with a mixed distribution of primary kidney disease and a higher prevalence of renovascular disease (8.3%). Approximately one third of patients in this cluster were listed for transplantation (35.2%).
Cluster 3 comprised the oldest phenotype (73.6 [7.9] years) and was characterized by the lowest hemoglobin and albumin levels (91.1 [9.5] and 31.6 [6.4] g/L). This cluster had the highest prevalence of renovascular disease (11.7%) and the lowest transplant-listing rate (11.5%). Sex distribution was similar across clusters. Differences in age, ethnicity, primary kidney disease, and transplant-listing status were statistically significant across clusters (all P < 0.001).
Cluster Stability
Clustering results were highly stable. Across 45 pairwise tcomparisons of ten repeated clustering runs, the mean ARI was 0.71, indicating substantial agreement in cluster assignments.
Mortality Differences between PD and Hemodialysis in the Overall Cohort
Mortality was analyzed using a competing-risks framework, with kidney transplantation treated as the competing event. sHRs were estimated using pseudo-value regression at prespecified time points of 1 and 5 years, adjusted for age, sex, ethnicity, primary kidney disease, hemoglobin, serum albumin, and phenotype membership.
Across the overall cohort, initiation of hemodialysis was associated with a higher cumulative incidence of death compared with PD at both time points (Tables 3 and 4). At 1 year, hemodialysis was associated with a substantially higher mortality risk (adjusted sHR, 1.85; 95% CI, 1.75 to 1.96; P < 0.001). At 5 years, this association persisted but was attenuated in magnitude (adjusted sHR, 1.47; 95% CI, 1.43 to 1.51; P < 0.001).
Table 3.
Competing risk regression across the whole cohort at 1 year time point
| Variable | Adjusted sHR | 95% CI | P Value |
|---|---|---|---|
| Dialysis modality (hemodialysis versus PD) | 1.85 | 1.75 to 1.96 | <0.001 |
| Cluster (phenotype) | |||
| Cluster 2 (versus cluster 1) | 1.02 | 0.94 to 1.11 | 0.63 |
| Cluster 3 (versus cluster 1) | 1.00 | 0.92 to 1.09 | 0.94 |
| Primary kidney disease (versus diabetic kidney disease) | |||
| GN | 0.69 | 0.64 to 0.75 | <0.001 |
| Hypertensive nephropathy | 0.85 | 0.77 to 0.93 | <0.001 |
| Other causes | 1.41 | 1.33 to 1.49 | <0.001 |
| Polycystic kidney disease | 0.52 | 0.44 to 0.61 | <0.001 |
| Pyelonephritis | 0.95 | 0.87 to 1.04 | 0.30 |
| Renovascular disease | 1.32 | 1.23 to 1.42 | <0.001 |
| Unknown | 1.57 | 1.49 to 1.65 | <0.001 |
| Ethnicity (versus White) | |||
| Asian | 0.63 | 0.58 to 0.68 | <0.001 |
| Black | 0.55 | 0.49 to 0.62 | <0.001 |
| Mixed | 0.91 | 0.74 to 1.13 | 0.41 |
| Other | 0.61 | 0.49 to 0.77 | <0.001 |
| Unknown | 1.69 | 1.61 to 1.77 | <0.001 |
| Male sex | 0.99 | 0.96 to 1.03 | 0.69 |
| Hemoglobin (per g/L) | 0.994 | 0.992 to 0.996 | <0.001 |
| Albumin (per g/L) | 0.92 | 0.92 to 0.92 | <0.001 |
| Age (per yr) | 1.03 | 1.03 to 1.03 | <0.001 |
CI, confidence interval; PD, peritoneal dialysis; sHR, subdistribution hazard ratio.
Table 4.
Competing risk regression across the whole cohort at 5-year time point
| Variable | Adjusted sHR | 95% CI | P Value |
|---|---|---|---|
| Dialysis modality (hemodialysis versus peritoneal dialysis) | 1.47 | 1.43 to 1.51 | <0.001 |
| Cluster (phenotype) | |||
| Cluster 2 (versus cluster 1) | 0.87 | 0.83 to 0.92 | <0.001 |
| Cluster 3 (versus cluster 1) | 0.87 | 0.83 to 0.92 | <0.001 |
| Primary kidney disease (versus diabetic kidney disease) | |||
| GN | 0.48 | 0.46 to 0.50 | <0.001 |
| Hypertensive nephropathy | 0.67 | 0.64 to 0.71 | <0.001 |
| Other causes | 0.82 | 0.79 to 0.85 | <0.001 |
| Polycystic kidney disease | 0.36 | 0.33 to 0.39 | <0.001 |
| Pyelonephritis | 0.63 | 0.60 to 0.67 | <0.001 |
| Renovascular disease | 0.93 | 0.88 to 0.98 | 0.004 |
| Unknown | 0.89 | 0.86 to 0.92 | <0.001 |
| Ethnicity (versus White) | |||
| Asian | 0.68 | 0.65 to 0.70 | <0.001 |
| Black | 0.57 | 0.53 to 0.60 | <0.001 |
| Mixed | 0.79 | 0.69 to 0.90 | <0.001 |
| Other | 0.66 | 0.59 to 0.74 | <0.001 |
| Unknown | 1.29 | 1.24 to 1.33 | <0.001 |
| Male sex | 1.01 | 0.98 to 1.03 | 0.46 |
| Hemoglobin (per g/L) | 0.996 | 0.995 to 0.997 | <0.001 |
| Albumin (per g/L) | 0.95 | 0.95 to 0.96 | <0.001 |
| Age (per yr) | 1.05 | 1.05 to 1.05 | <0.001 |
CI, confidence interval; sHR, subdistribution hazard ratio.
At 1 year, after adjustment for dialysis modality and baseline covariates, mortality risk was similar across phenotypes. By 5 years, a clear phenotype effect had emerged, with clusters 2 and 3 demonstrating significantly lower adjusted mortality than cluster 1, independent of dialysis modality (adjusted sHR, 0.87 for both clusters; P < 0.001). Older age, lower serum albumin, and lower hemoglobin were independently associated with higher mortality at both time points. Compared with White ethnicity, Asian and Black ethnicity were associated with lower adjusted mortality, whereas unknown ethnicity was associated with higher mortality. Associations with primary kidney disease were directionally consistent across time points, with lower mortality observed for GN, hypertensive nephropathy, polycystic kidney disease, and pyelonephritis compared with diabetic kidney disease. Cumulative incidence functions for mortality in the overall cohort and stratified by phenotype are shown in Figures 1–4.
Figure 1.
Cumulative incidence of death by dialysis modality from dialysis initiation to 5 years across the whole cohort. Curves represent the CIF for all-cause mortality among incident dialysis patients initiating PD or hemodialysis, with kidney transplantation treated as a competing event. Estimates are shown for the overall cohort over the first 5 years of follow-up. CIF, cumulative incidence function; PD, peritoneal dialysis.
Figure 4.
Cumulative incidence of death by dialysis modality from dialysis initiation to 5 years across cluster 3. Curves represent the CIF for all-cause mortality among incident dialysis patients initiating PD or hemodialysis, with kidney transplantation treated as a competing event. Estimates are shown for cluster 3 over the first 5 years of follow-up.
Figure 2.
Cumulative incidence of death by dialysis modality from dialysis initiation to 5 years across cluster 1. Curves represent the CIF for all-cause mortality among incident dialysis patients initiating PD or hemodialysis, with kidney transplantation treated as a competing event. Estimates are shown for cluster 1 over the first 5 years of follow-up.
Figure 3.
Cumulative incidence of death by dialysis modality from dialysis initiation to 5 years across cluster 2. Curves represent the CIF for all-cause mortality among incident dialysis patients initiating PD or haemodialysis, with kidney transplantation treated as a competing event. Estimates are shown for the cluster 2 over the first 5 years of follow-up.
Stratified Phenotype-Specific Analyses
Adjusted sHRs comparing hemodialysis with PD were estimated within each phenotype using stratified competing-risks models (Table 5). At 1 year, hemodialysis was associated with higher mortality across all three phenotypes, with adjusted sHRs of 2.29 (95% CI, 1.87 to 2.80) in cluster 1, 2.15 (95% CI, 1.95 to 2.38) in cluster 2, and 1.55 (95% CI, 1.44 to 1.68) in cluster 3.
Table 5.
Adjusted 1- and 5-year subdistribution hazard ratios for peritoneal dialysis versus hemodialysis in the overall cohort and by cluster
| Group | sHR for hemodialysis versus PD at 1 yr | sHR for hemodialysis versus PD at 5 yr |
|---|---|---|
| Cluster 1 | 2.29 (95% CI, 1.87 to 2.80; P < 0.001) | 2.42 (95% CI, 2.23 to 2.63; P < 0.001) |
| Cluster 2 | 2.15 (95% CI, 1.95 to 2.38; P < 0.001) | 1.55 (95% CI, 1.48 to 1.61; P < 0.001) |
| Cluster 3 | 1.55 (95% CI, 1.44 to 1.68; P < 0.001) | 1.14 (95% CI, 1.11 to 1.19; P < 0.001) |
CI, confidence interval; PD, peritoneal dialysis; sHR, subdistribution hazard ratio.
At 5 years, hemodialysis remained associated with higher mortality in all phenotypes, although the magnitude of association varied across clusters. The adjusted sHRs were 2.42 (95% CI, 2.23 to 2.63) in cluster 1, 1.55 (95% CI, 1.48 to 1.61) in cluster 2, and 1.14 (95% CI, 1.11 to 1.19) in cluster 3. These findings indicate that the relative mortality difference between modalities was greatest in the youngest phenotype and progressively attenuated with increasing age.
Model Discrimination
Model discrimination was assessed using the cross-validated concordance index (C-index). The C-index was 0.68, indicating moderate discrimination. As the objective of the analysis was estimation of adjusted associations rather than development of a prediction model, the C-index was used to characterize overall model performance.
Time to Transplantation across Phenotypes
In a complementary competing-risks analysis with kidney transplantation as the event of interest and death as the competing event, time to transplantation was evaluated at 5 years after dialysis initiation. After multivariable adjustment, there were no statistically significant differences in transplantation incidence across phenotypes. Compared with cluster 1, the adjusted sHR for transplantation was 1.04 for cluster 2 (P = 0.28) and 1.06 for cluster 3 (P = 0.13). These findings suggest that differential access to transplantation is unlikely to account for the observed phenotype-specific differences in mortality.
Health System Context
According to the ISN GKHA 2023, the United Kingdom has a higher burden of kidney failure than Western Europe, with an incidence of 151 per million population compared with 135 per million population and a prevalence of 1293 compared with 1034 per million population (Table 6). Despite this, dialysis infrastructure is more limited in the United Kingdom, with 1.0 hemodialysis and 1.0 PD center per million population, compared with 7.7 and 2.7 centers per million population, respectively, across Western Europe.
Table 6.
Comparison between the UK and the Western Europe data
| Metric | United Kingdom | Western Europe |
|---|---|---|
| Prevalence of treated kidney failure | 1293 pmp (↑35.3% since 2019) | 1034 pmp (↑5.6% since 2019) |
| Incidence of kidney failure | 151 pmp (↑30.5% since 2019) | 135 pmp (↑5.5% since 2019) |
| Annual cost—In-center hemodialysis | $45,169.7 (↓9.1% since 2019) | $66,351 (↑10.5% since 2019) |
| CAPD | $22,286.8 (↓29.3% since 2019) | $35,464.9 (↓18.8% since 2019) |
| Cost—Kidney transplant (first yr) | $14,011.8 (↓49.9% since 2019) | $74,089.2 (↑17.2% since 2019) |
| Annual cost—Kidney transplant (maintenance, later yr) | $4547 (↓13.0% since 2019) | $14,479.7 (↑2.2% since 2019) |
| Dialysis center availability (PMP) | Hemodialysis: 1.0 PMP Peritoneal dialysis: 1.0 PMP Transplant centers: 0.34 PMP |
Hemodialysis: 7.7 PMP Peritoneal dialysis: 2.7 PMP Transplant centers: 0.55 PMP |
CAPD, annual costs–PD; PMP, per million population.
Annual dialysis costs were lower in the United Kingdom than the Western European median (in-center hemodialysis US $45,170 versus US $66,351; PD US $22,287 versus US $35,465). First-year kidney transplantation costs were also substantially lower in the UK (US $14,012 versus US $74,089). These cost comparisons are intended to contextualize dialysis delivery across health systems rather than to provide a formal economic evaluation.
Discussion
In this large contemporary UKRR cohort of 96,809 incident dialysis patients, initiation of PD was associated with a lower cumulative incidence of death compared with hemodialysis, although the magnitude and persistence of this association varied by patient phenotype and over time. Using unsupervised k-prototypes clustering, we identified three reproducible phenotypes capturing major sources of demographic, biochemical, and disease-related heterogeneity. Time-specific competing-risk analyses demonstrated that PD was associated with lower mortality at 1 year across all phenotypes, while at 5 years this association persisted but attenuated with increasing age, remaining strongest in the youngest phenotype.
Earlier UK studies comparing survival between PD and hemodialysis largely examined cohorts from the 1990s and early 2000s and reported heterogeneous findings. Single-center and regional analyses by McKane et al. and Traynor et al. observed broadly similar survival between modalities.11,12 Nitsch et al., analyzing UKRR data from 1997 to 2005, reported modest early differences favoring PD that attenuated after approximately 2 years.13 These studies predated widespread adoption of low-glucose degradation product solutions, automated cyclers, neutral-pH fluids, structured home-therapy training, and improvements in peritoneal technology and supportive care.14 International registry analyses and systematic reviews have similarly reported early survival differences favoring PD that diminish over time, including European and United States cohorts and the most recent Cochrane review of observational data.3,15,16 Collectively, these findings highlight persistent uncertainty regarding the durability of modality-associated survival differences and underscore the need for contemporary, phenotype-informed analyses.
Our analysis addresses this gap through data-driven phenotyping rather than predefined clinical subgroups. The resulting phenotypes aligned with clinically recognizable profiles. The youngest phenotype was characterized by higher transplant listing rates and greater representation of GN and polycystic kidney disease. The intermediate phenotype demonstrated preserved nutritional status with mixed disease etiology, while the oldest phenotype was marked by frailty, lower hemoglobin, and albumin and a higher burden of renovascular disease. High cluster stability supports the reproducibility of these phenotypes.
Several factors may plausibly contribute to the heterogeneity in modality-associated mortality observed across phenotypes. The stronger relative association observed in the youngest phenotype may reflect longer dialysis exposure and lower competing mortality, allowing modality-related differences to accrue over time. By contrast, the more attenuated association seen in older and frailer phenotypes may reflect higher competing risks of death, shorter effective dialysis duration, and reduced physiologic reserve, which may limit the long-term impact of dialysis modality choice. Additional factors, such as differences in residual kidney function, tolerance of intradialytic hemodynamic stress, and treatment burden, may also contribute to these phenotype-specific patterns. These potential mechanisms cannot be directly evaluated within the present observational analysis and should be interpreted as hypotheses rather than causal explanations.
Placing these findings in a broader health system context is important. According to the ISN GKHA, the United Kingdom has a higher burden of kidney failure than Western Europe, with an incidence of 151 pmp (per million population) compared with 135 pmp and a prevalence of 1293 pmp compared with 1034 pmp, while having among the lowest dialysis center densities (1.0 hemodialysis and 1.0 PD centers pmp).7 Despite this, annual dialysis costs in the United Kingdom are substantially lower than Western European medians (in-center hemodialysis US $45,170 versus US $66,351; PD US $22,287 versus US $35,465), and first-year kidney transplantation costs are markedly lower (US $14,012 versus US $74,089).7 Between 2019 and 2023, PD costs decreased in both the United Kingdom and Western Europe, although PD uptake in the United Kingdom has remained persistently low.
These findings should be interpreted within the constraints of an observational study. Dialysis modality choice is influenced by a complex interplay of factors not fully captured in the registry, including clinician counseling practices, center expertise, availability of assisted PD and home-therapy training, and patient-level considerations such as home environment, social support, and personal preference. Although we adjusted for key demographic, laboratory, and disease-related variables and used phenotype-based stratification to reduce clinical heterogeneity, residual confounding cannot be excluded, and the observed associations should not be interpreted as causal.
Temporal changes in dialysis practice represent a further limitation. The study period spans more than a decade (2007–2021), during which there have been substantial changes in dialysis technology, clinical practice, patient selection, and supportive care, including improvements in vascular access management, dialysis adequacy, and multidisciplinary support. Although the UKRR provides consistent national coverage across this period, unmeasured temporal heterogeneity may have influenced both modality choice and outcomes. Our analyses did not explicitly stratify by calendar era, and therefore, the reported associations should be interpreted as reflecting average effects across contemporary UK practice rather than effects within a single uniform treatment epoch.
In addition, more contemporary hemodialysis modalities, such as online hemodiafiltration, home hemodialysis, or nocturnal hemodialysis, were not reliably captured in the UKRR during much of the study period and could not be evaluated separately. As uptake of hemodiafiltration has increased in parts of Europe and globally, outcomes associated with modern hemodialysis practice may differ from those observed here. Accordingly, our findings should be interpreted in the context of the dialysis modalities as recorded in the United Kingdom during the study period.
This study has strengths including a large national cohort, comprehensive capture of modality transitions and transplantation, a contemporary timeframe, and reproducible clustering methodology. Nevertheless, several clinically relevant covariates—such as frailty, detailed cardiovascular comorbidity, residual kidney function, dialysis adequacy, vascular access type, center-level practice patterns, socioeconomic factors, and patient preference—were unavailable or incompletely captured. Time-varying factors such as modality switching, intercurrent illness, and evolving treatment strategies were not modeled. As a result, unmeasured and residual confounding cannot be excluded.
While PD was associated with lower system-level costs compared with in-center hemodialysis, these analyses are descriptive and do not constitute a formal cost-effectiveness evaluation. The UKRR does not capture patient-level cost data, quality-of-life measures, or utilities required to estimate quality-adjusted life years or incremental cost-effectiveness ratios. However, the phenotype-based framework presented here may support future United Kingdom–specific economic evaluations by identifying patient subgroups in whom survival differences between modalities are greatest.
In conclusion, initiation of PD was associated with a lower cumulative incidence of death compared with hemodialysis in contemporary UK practice, with time-dependent variation across patient phenotypes. When considered alongside international health system data, these findings highlight a persistent evidence-practice gap and support consideration of more personalized, phenotype-informed modality selection and alignment of dialysis capacity with contemporary patterns of care.
Acknowledgment
Tibor Fülöp is a current employee of the United States Veterans Health Administration. However, the opinions and views expressed in this paper are the Authors' own and do not represent the official views or policies of the United States Veteran Health Administrations.
Disclosures
Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/CJN/C718.
Author Contributions
Conceptualization: Hatem Ali, Rizwan Hamer.
Data curation: Hatem Ali, Anna Maria Casula, Andre Paola Ortega Alban.
Formal analysis: Hatem Ali.
Methodology: Hatem Ali, Rizwan Hamer.
Validation: Hatem Ali.
Writing – original draft: Hatem Ali.
Writing – review & editing: Samar Abd ElHafeez, Tibor Fülöp.
Funding
None.
Declarative Statements
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.
Data Availability Statements
Original data generated for the study will be made available upon reasonable request to the corresponding author. Data Type: Observational Data. Reason for Restricted Access: The data that support the findings of this study are derived from the UKRR. Restrictions apply to the availability of these data, which were used under a data-sharing agreement (DSA125) for the current study and are therefore not publicly available. De-identified data may be made available upon reasonable request to the UKRR (https://ukkidney.org/audit-research/renal-registry) subject to approval of a data-sharing application and compliance with governance requirements. The code used for clustering and survival analyses is available from the authors on request.
References
- 1.Jha V Garcia-Garcia G Iseki K, et al. Chronic kidney disease: global dimension and perspectives. Lancet. 2013;382(9888):260–272. doi: 10.1016/S0140-6736(13)60687-X [DOI] [PubMed] [Google Scholar]
- 2.Johansen KL. Does peritoneal dialysis offer an early survival advantage over hemodialysis? Semin Dial. 2007;20(2):120–124. doi: 10.1111/j.1525-139X.2007.00280.x [DOI] [Google Scholar]
- 3.Ethier I Hayat A Pei J, et al. Peritoneal dialysis versus haemodialysis for people commencing dialysis. Cochrane Database Syst Rev. 2024;6(6):CD013800. doi: 10.1002/14651858.CD013800.pub2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.UK Renal Registry. UK Renal Registry 24th Annual Report. UK Renal Registry; 2021. https://www.ukkidney.org/audit-research/annual-report. [Google Scholar]
- 5.Wagner M Ansell D Kent DM, et al. Predicting mortality in incident dialysis patients. Am J Kidney Dis. 2011;57(6):894–902. doi: 10.1053/j.ajkd.2011.01.012 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Huang Z. Extensions to the k-means algorithm for clustering large data sets with categorical values. Data Mining Knowledge Discov. 1998;2(3):283–304. doi: 10.1023/a:1009769707641 [DOI] [Google Scholar]
- 7.International Society of Nephrology. Global Kidney Health Atlas 2023 (ISN-GKHA). Kidney Int Suppl. 2024;13(1) https://www.theisn.org/in-action/research/global-kidney-health-atlas/. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Kaufman L, Rousseeuw PJ. Finding Groups in Data: An Introduction to Cluster Analysis. Wiley; 1990. [Google Scholar]
- 9.Hubert L, Arabie P. Comparing partitions. J Classification. 1985;2(1):193–218. doi: 10.1007/bf01908075 [DOI] [Google Scholar]
- 10.Pedregosa F Varoquaux G Gramfort A, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. Nephrol Dial Transplant. 2011;12(85):2825–2830. [Google Scholar]
- 11.McKane WR Cameron JS Howie AJ, et al. Comparative survival and clinical outcomes of peritoneal dialysis and haemodialysis: a single-centre study in Sheffield. Nephrol Dial Transplant. 2002;17(11):1874–1880. doi: 10.1093/ndt/17.11.1946 [DOI] [Google Scholar]
- 12.Traynor JP, Simpson K, Geddes CC, Deighan CJ, Fox JG. Early initiation of dialysis fails to prolong survival in patients with end-stage renal failure. J Am Soc Nephrol. 2002;13(8):2125–2132. doi: 10.1097/01.ASN.0000025294.40179.E8 [DOI] [PubMed] [Google Scholar]
- 13.Nitsch D Steenkamp R Tomson CRV, et al. Chapter 8: Survival and cause of death in UK adult patients on renal replacement therapy in 2010: national and centre-specific analyses. Nephron Clin Pract. 2012;120(Suppl 1):c117–c153 [DOI] [PubMed] [Google Scholar]
- 14.Findlay MD Jardine AG McQuarrie EP, et al. Trends in patient survival on peritoneal dialysis and hemodialysis in Scotland: a registry-based analysis. Nephron. 2018;138(1):31–41. doi: 10.1159/000481672 [DOI] [Google Scholar]
- 15.van de Luijtgaarden MW Jager KJ Segelmark M, et al. Trends in dialysis modality choice and related patient survival in the ERA-EDTA registry over a 20-year period. Nephrol Dial Transplant. 2016;31(1):120–128. doi: 10.1093/ndt/gfv295 [DOI] [PubMed] [Google Scholar]
- 16.Mehrotra R, Chiu YW, Kalantar-Zadeh K, Bargman J, Vonesh E. Similar outcomes with hemodialysis and peritoneal dialysis in patients with end-stage renal disease. Arch Intern Med. 2011;171(2):110–118. doi: 10.1001/archinternmed.2010.352 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
Original data generated for the study will be made available upon reasonable request to the corresponding author. Data Type: Observational Data. Reason for Restricted Access: The data that support the findings of this study are derived from the UKRR. Restrictions apply to the availability of these data, which were used under a data-sharing agreement (DSA125) for the current study and are therefore not publicly available. De-identified data may be made available upon reasonable request to the UKRR (https://ukkidney.org/audit-research/renal-registry) subject to approval of a data-sharing application and compliance with governance requirements. The code used for clustering and survival analyses is available from the authors on request.