Comprehensive Summary
In this study, researchers used unsupervised phenotypic clustering to identify two distinct clusters of 319 ICM (ischemic cardiomyopathy) patients. Cardiomagnetic resonance (CMR)-derived data, such as left-ventricular (LV) volume and scarring, was implemented into the KAMILA clustering algorithm to identify two distinct phenotypes: patients with better cardiac function (Cluster 1), and patients with more advanced stages of disease (Cluster 2). These clusters varied significantly in variables such as LV ejection fraction, which Cluster 1 was found to have a higher value (43.0 ± 6.3% vs. 28.0 ± 7.3%) and midwall fibrosis, which Cluster 2 was found to have a higher value (6.8% vs. 17.3%, p = 0.007). Additionally, researchers utilized a SHAP summary plot to demonstrate how much influence each variable had on the model’s cluster predictions, with the ischemic scar (% of LV mass) having the heaviest contribution. Researchers followed-up with each patient to assess for any of the five predefined composite outcomes (cardiovascular death, aborted sudden cardiac death, ICD therapy, heart failure hospitalization, and LVAD implantation or heart transplant). Notably, patients in Cluster 2 were almost four times as likely to experience one of the composite outcomes than patients in Cluster 1 (HR = 3.96, 95% CI: 2.02–7.76, p < 0.001).
Outcomes and Implications
Despite the use of CMR and established risk models to analyze cardiac function and negative outcome risk, ICM patients are still experiencing variable clinical outcomes. Using unsupervised clustering to identify different ICM phenotypes could help clinicians identify patients with a higher hazard ratio and provide more aggressive treatment and monitoring, while patients with a lower risk can receive less intensive treatment. However, it should be acknowledged that this study was performed on a small sample size, was not externally validated, and was not statistically compared to pre-existing risk models. Once validated with a larger cohort and refined with additional data types, unsupervised clustering can be a helpful tool for clinicians to personalize their patient’s treatment and provide better outcomes.