Published online Sep 18, 2026. doi: 10.5500/wjt.121821
Revised: June 21, 2026
Accepted: June 29, 2026
Published online: September 18, 2026
Processing time: 150 Days and 0.3 Hours
Unsupervised machine learning identifies clinically meaningful subgroups across biomedical domains, yet remains underutilised in transplantation. Existing machine perfusion (MP) machine learning research relies on small cohorts with supervised models of variable performance, rarely investigating multivariate phenotypes integrating donor-recipient characteristics. Unsupervised methods can uncover latent phenotypes in high-dimensional data, revealing favourable-outcome groups and high-risk profiles overlooked by current scoring systems. Applying these methods to United Kingdom MP data may yield novel insights from the underexplored National Health Service Blood and Transplant (NHSBT) dataset. We hypothesised that MP cases occupy preferential positions within data-derived clinical phenotypes from unsupervised clustering, and these phenotypes retain structural stability and outcome stratification in the MP subgroup.
To evaluate whether MP cases are non-randomly distributed across unsupervised clinical clusters and whether MP status modifies cluster structure or outcome stratification.
We analysed the standard NHSBT kidney dataset comprising all United Kingdom adult kidney transplants from 2014 to 2024. The full cohort had 15904 cases and a machine-perfused subgroup of 544. Dimensionality reduction was conducted using a mixed-type principal component analysis embedding approach (FAMD-like) to generate a shared latent space for all features. Gaussian mixture model was fitted to the full cohort to define latent phenotypes, and then transferred to the MP subgroup. We performed descriptive post hoc characterisation of the cluster and assessed graft and patient survival outcomes.
Six distinct clinical phenotypes were identified with moderate-to-good stability. MP cases showed non-random distribution, with notable enrichment in cluster 6 (40% of MP patients) and differential perfusion-type associations (hypothermic MP: Clusters 3, 5, 6; normothermic MP: Clusters 1, 4). When applied to the MP subgroup, cluster definitions remained stable, though with modestly reduced assignment confidence (mean entropy 0.815 vs 0.783), indicating MP cases are embedded within the general population structure. Clusters retained outcome stratification in the MP subgroup, though outcome patterns diverged in specific phenotypes: Cluster 2 showed concentrated early graft failures (all within 2.07 years), while cluster 4 had lower patient survival despite stability in the full cohort. These findings suggest MP status may interact with underlying clinical phenotypes to modify outcome trajectories.
MP cases showed non-random cluster distribution and perfusion-type associations. Cluster structure remained stable with retained outcome stratification in the MP subgroup, though outcome patterns differed in specific phenotypes.
Core Tip: Using unsupervised clustering, we identified latent clinical phenotypes that demonstrated a non-random distribution of machine perfusion (MP) across the population. MP cases were notably concentrated within specific clusters, while the overall cluster structure remained mostly stable when focusing solely on MP recipients. This indicates that MP is integrated within the wider phenotypic landscape rather than creating distinct subgroups. Importantly, these phenotypes continued to hold outcome patterns within the MP cohort, although some varied among certain clusters. These findings suggest that MP interacts with the underlying profiles of recipients and donors, potentially altering outcome trajectories depending on the specific phenotype.