Cardiovascular

Exploratory Mixed-Data Phenomapping and Risk-Enrichment Frameworks for Long-Term Mortality After Acute Coronary Syndrome: A Single-Center Cohort Study.

TL;DR

Outcome-independent phenomapping identified exploratory, method-dependent patterns rather than ordered biological risk classes, and post hoc additive frameworks were internally associated with mortality but should not be regarded as validated prediction or treatment-directing tools.

Key Findings

Three Gower-PAM phenotypes were identified in ACS patients with significantly different crude mortality proportions across groups.

  • Three phenotypes were selected based on prespecified outcome-independent silhouette, minimum-size, and bootstrap-stability criteria, with an average silhouette width of 0.417.
  • Crude mortality proportions were 12.1% in phenotype A, 4.2% in phenotype B, and 19.2% in phenotype C.
  • Global log-rank p < 0.001 across the three phenotypes.
  • 3491 patients were included in the Gower-PAM phenomapping after exclusions for incomplete clustering variables.

The phenotype factor improved the primary age-inclusive model, but only phenotype B differed significantly from phenotype A after adjustment.

  • The phenotype factor had a joint p < 0.001 in the primary age-inclusive model.
  • Only phenotype B, not phenotype C, differed from phenotype A after adjustment.
  • The joint phenotype effect was attenuated after expanded clinical adjustment (p = 0.673).
  • Agreement with k-means sensitivity analyses was low, indicating method dependence of the identified patterns.

All three additive risk frameworks (AAR-3, CARE-4, ACEF) were associated with first-year mortality after ACS with hazard ratios ranging from 1.78 to 2.66.

  • First-year adjusted HR was 1.78 per 1-point increase in AAR-3 (95% CI, 1.46–2.19).
  • First-year adjusted HR was 2.08 per 1-point increase in CARE-4 (95% CI, 1.76–2.46).
  • First-year adjusted HR was 2.66 per 1-unit increase in ACEF (95% CI, 2.14–3.30).
  • These associations were observed in a common complete-case cohort of 3492 patients compared against an age-inclusive clinical reference model.

All three additive risk frameworks remained associated with mortality beyond the first year in landmark analyses.

  • The >365-day landmark HR for AAR-3 was 1.56 (95% CI, 1.32–1.85).
  • The >365-day landmark HR for CARE-4 was 1.66 (95% CI, 1.44–1.91).
  • The >365-day landmark HR for ACEF was 2.39 (95% CI, 1.95–2.92).
  • These landmark analyses indicate that the association with mortality persisted beyond the acute phase.

Full-follow-up C-indices for the three additive frameworks ranged from 0.793 to 0.815, all exceeding the age-inclusive reference model's C-index of 0.770.

  • C-index was 0.793 (95% CI, 0.772–0.814) with AAR-3.
  • C-index was 0.815 (95% CI, 0.795–0.834) with CARE-4.
  • C-index was 0.812 (95% CI, 0.792–0.833) with ACEF.
  • The age-inclusive reference model had a C-index of 0.770 (95% CI, 0.747–0.792).
  • None of the three frameworks materially improved the model that already contained age, hemoglobin, eGFR, and LVEF as continuous variables.

During a median follow-up of 3.35 years, 407 deaths occurred among the 3561 patients in the survival cohort.

  • 3801 consecutive patients were hospitalized with ACS during 2017–2024; 240 were excluded for lacking mortality status or valid follow-up, leaving 3561 in the survival cohort.
  • Median follow-up was 3.35 years.
  • 407 total deaths occurred during follow-up.
  • 70 additional patients were excluded from the Gower-PAM clustering step due to incomplete clustering variables.

Outcome-independent phenomapping produced exploratory, method-dependent patterns rather than stable biological risk classes in ACS patients.

  • Agreement between Gower-PAM and k-means sensitivity analyses was low, indicating instability of identified clusters.
  • The joint phenotype effect was no longer significant after expanded clinical adjustment (p = 0.673).
  • Authors conclude that the phenotypes should be regarded as 'exploratory, method-dependent patterns rather than ordered biological risk classes.'
  • Prospective multicenter external validation was stated to be required before clinical application.

None of the three additive risk frameworks materially improved discrimination beyond a model containing age, hemoglobin, eGFR, and LVEF as continuous variables.

  • AAR-3, CARE-4, and ACEF each showed higher C-indices than the simple age-inclusive reference model.
  • However, none materially improved the model already containing age, hemoglobin, eGFR, and LVEF continuously.
  • Authors caution that these frameworks 'should not be regarded as validated prediction or treatment-directing tools.'
  • The study was a single-center cohort without external validation.

What This Means

This research examined whether grouping heart attack (acute coronary syndrome, or ACS) patients into distinct biological 'types' based on their baseline characteristics, and whether simple scoring systems, could predict which patients were most likely to die over the following years. The study followed 3,561 patients admitted to a single hospital between 2017 and 2024, with a median follow-up of about 3.35 years, during which 407 patients died. Using a statistical clustering technique, researchers identified three patient groups ('phenotypes') with notably different death rates—about 4%, 12%, and 19%—but when they accounted for known clinical factors like age, kidney function, blood count, and heart function, most of the apparent differences between groups disappeared, and the groupings changed depending on which clustering method was used. Three scoring systems—AAR-3, CARE-4, and ACEF—were each linked to higher mortality risk both in the first year and in subsequent years after a heart attack, and all three performed better than a simple age-based reference model. However, none of the three scores added meaningful predictive value beyond a model that already included age, hemoglobin, kidney function (eGFR), and heart pumping function (LVEF) measured as continuous variables. This suggests that the information captured by these scores largely overlaps with these four basic clinical measurements. This research suggests that while clustering ACS patients into groups and using additive risk scores can reveal statistical patterns, these approaches have important limitations: the patient groupings are sensitive to the method used and may not reflect true biological differences, and the scoring systems—though associated with outcomes—have not been externally validated and should not yet be used to guide clinical decisions. The authors emphasize that multicenter prospective studies are needed before these tools can be applied in real-world patient care.

Have a question about this study?

Citation

&#xc7;elik A, Kocaba&#x15f; U, K&#x131;r&#x131;&#x15f; T, Af&#x15f;in A, Erdem H, U&#xe7;ak E, et al.. (2026). Exploratory Mixed-Data Phenomapping and Risk-Enrichment Frameworks for Long-Term Mortality After Acute Coronary Syndrome: A Single-Center Cohort Study.. Medicina (Kaunas, Lithuania). https://doi.org/10.3390/medicina62081586