Not all pci patients share the same risk trajectory: unsupervised clustering identifies three clinical phenotypes with divergent ischaemia–bleeding–mortality profiles

Background Patients undergoing PCI are clinically heterogeneous, and single-variable stratification often fails to capture distinct risk trajectories. Data-driven unsupervised phenotyping in real-world cohorts may identify clinically interpretable subgroups to support refined risk assessment and follow-up. Purpose To identify clinical phenotypes among PCI-treated patients using unsupervised clustering and to evaluate phenotype-specific differences in ischaemic, bleeding, and mortality-related outcomes. Methods Data were derived from a real-world coronary artery disease specialty database of PCI-treated patients from 82 hospitals, with an independent single-center PCI cohort for external validation. We performed unsupervised phenotyping using the k-prototypes algorithm. Kaplan–Meier curves were used to compare cumulative event risks across phenotypes. For the primary outcome, multivariable Cox proportional hazards models were fitted to estimate adjusted hazard ratios (HRs) with 95% confidence intervals (CIs), using Cluster 2 as the reference. The primary endpoint was the composite NACE, defined as the occurrence of any of the following: cardiovascular death (CD), myocardial infarction (MI), intracranial haemorrhage, ischaemic stroke, or Bleeding Academic Research Consortium type 3 bleeding (BARC3). Secondary endpoints were the individual components of these composites. Results A total of 197,387 PCI-treated patients were classified into three clinically interpretable phenotypes: Cluster 1 (n=7,002), Cluster 2 (n=107,054), and Cluster 3 (n=83,331). Cluster 1 was an acute high-risk, event-driven phenotype with a high STEMI prevalence (45.4%). Cluster 3 was an older, female-predominant chronic-burden phenotype (age 63.62±9.79 years; female 57.1%), distinct from Cluster 2 (STEMI 12.9%; female 34.5%) (all P<0.001).Over 1-year follow-up, NACE Kaplan–Meier curves differed significantly among phenotypes (P<0.001). In multivariable Cox models, both Cluster 1 and Cluster 3 had higher risk of NACE compared with Cluster 2 (Cluster 1: HR 1.138, 95% CI 1.042–1.242, P=0.004; Cluster 3: HR 1.122, 95% CI 1.079–1.167, P<0.001). Cluster 3 also had increased risk of BARC type 3 bleeding (HR 1.269, 95% CI 1.172–1.375, P<0.001). All-cause mortality was significantly higher in Cluster 1 and Cluster 3 (Cluster 1: HR 1.371, 95% CI 1.265–1.485; Cluster 3: HR 1.270, 95% CI 1.217–1.325; both P<0.001). In contrast, MACCE was lower in Cluster 3 (HR 0.894, 95% CI 0.867–0.923, P<0.001), suggesting that composite endpoints may be influenced by differences in event composition across phenotypes. External validation showed consistent directions of phenotype–outcome associations. Conclusion(s) Unsupervised clustering identified three PCI phenotypes with distinct ischaemia–bleeding–mortality profiles, enabling finer risk stratification and underscoring the need to interpret composite endpoints alongside key individual outcomes.For image description, please refer to the figure legend and surrounding text.1-Year Clinical OutcomesFor image description, please refer to the figure legend and surrounding text.