Machine learning–based decision trees classified chronic HF patients into low, intermediate, and high risk groups with HRs 3.3 and 4.3 for intermediate and high risk of HF readmission or death.
Does a machine learning-derived decision tree accurately stratify the 1-year risk of HF readmission or death in patients with chronic heart failure referred to Cardiology?
A machine learning-derived decision tree using 7 simple variables can effectively stratify the 1-year risk of death or HF readmission in chronic heart failure patients referred via e-consultancy.
Abstract Background Decision trees, obtained by machine learning algorithms, have demonstrated to be useful for the identification of clinical patterns with differential clinical profile. We tested whether this strategy could identify patients with chronic heart failure (HF) referred to Cardiology, by e-consultancy, with higher risk of HF readmission or dead in the first year. Methods Decision trees were obtained by Chi-square automatic interaction detection and classification and regression trees. The primary endpoint was HF readmission or death in the first year using the institutional records. All data were downloaded directly from the informatic system without manipulation from the investigators. Results We included 6,379 consecutive patients, referred to Cardiology consultancy between 2010-2021, with the previous diagnosis HF, mean age 77.6 (9.8) years and 49% women. The incidence of the composite primary endpoint was 16.5% (n=1,053 patients). The decision tree provided 11 clusters based on 7 variables: HF admission in the previous year, age, sex, prevalent neoplasm, delay to e-consultancy answer, sex, diabetes mellitus and previous coronary heart disease (figure 1); HF admission in the first year was the first discriminatory variable. The cluster 4 was the most prevalent (n=1,196; 18.8%) and had one of the lowest rates of the primary endpoint (6.7%), similar to cluster 5 (6.1%). Taking cluster 4 as the reference variable, the independent risk of HF readmission or death; clusters 5, 9 and 10 had no increased risk as compared to cluster 4. We collapsed the clusters in 3 categories according to the even rates: low risk, with clusters 4, 5, 9 and 10; intermediate risk with clusters 1, 3, 6,8 and 12; and high risk with clusters 2, 7 and 11. The Kaplan-Meyer curves (figure 2) depicted the clearly differentiated incidence that was statistically significant (log-rank test p0.01). The Cox regression analyses verified the independent risk of the intermediate risk group (HR: 3.3 95% CI 2.6-4.3; p0.01) and high-risk group (HR: 4.3 % CI 6.3-10.9; p0.01). No intra-group differences were observed in the low or intermediate risk categories; in contrast cluster 11 had significantly higher risk than clusters 2 and 7 in the high-risk categories. Conclusion a decision support system, obtained by machine learning algorithms, provided a simple classification of chronic HF patients based on 7 variables. The implementation of this algorithm when HF patients are referred to a Cardiology consultancy might help decision-making and provide actual risk stratification.Decision tree
Cordero et al. (2025) studied this question. Machine learning–based decision trees classified chronic HF patients into low, intermediate, and high risk groups with HRs 3.3 and 4.3 for intermediate and high risk of HF readmission or death.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: