Predicting mortality and re-hospitalization for heart failure: a machine-learning and cluster analysis on frailty and comorbidity

Background: Machine-learning techniques have been recently utilized to predict the probability of unfavorable outcomes among elderly patients suffering from heart failure (HF); yet none has integrated an assessment for frailty and comorbidity. This research seeks to determine which machine-learning-based phenogroups that incorporate frailty and comorbidity are most strongly correlated with death or readmission at hospital for HF within six months following discharge from hospital. Methods: In this single-center, prospective study of a tertiary care center, we included all patients aged 65 and older discharged for acute decompensated heart failure. Random forest analysis and a Cox multivariable regression were performed to determine the predictors of the composite endpoint. By k-means and hierarchical clustering, those predictors were utilized to phenomapping the cohort in four different clusters. Results: A total of 571 patients were included in the study. Cluster analysis identified four different clusters according to frailty, burden of comorbidities and BNP. As compared with Cluster 4, we found an increased 6-month risk of poor outcomes patients in Cluster 1 (very frail and comorbid; HR 3.53 [95% CI 2.30–5.39]), Cluster 2 (pre-frail with low levels of BNP; HR 2.59 [95% CI 1.66–4.07], and in Cluster 3 (pre-frail and comorbid with high levels of BNP; HR 3.75 [95% CI 2.25–6.27])). Conclusions: In older patients discharged for ADHF, the cluster analysis identified four distinct phenotypes according to frailty degree, comorbidity, and BNP levels. Further studies are warranted to validate these phenogroups and to guide an appropriate selection of personalized, model of care.