Exploring and Identifying Prognostic Phenotypes of Patients with Heart Failure Guided by Explainable Machine Learning

Abstract

Identifying patient prognostic phenotypes facilitates precision medicine. This study aimed to explore phenotypes of patients with heart failure (HF) corresponding to prognostic condition (risk of mortality) and identify the phenotype of new patients by machine learning (ML). A unsupervised ML was applied to explore phenotypes of patients in a derivation dataset (n = 562) based on their medical records. Thereafter, supervised ML models were trained on the derivation dataset to classify these identified phenotypes. Then, the trained classifiers were further validated on an independent validation dataset (n = 168). Finally, Shapley additive explanations were used to interpret decision making of phenotype classification. Three patient phenotypes corresponding to stratified mortality risk (high, low, and intermediate) were identified. Kaplan–Meier survival curves among the three phenotypes had significant difference (pairwise comparison p < 0.05). Hazard ratio of all-cause mortality between patients in phenotype 1 (n = 91; high risk) and phenotype 3 (n = 329; intermediate risk) was 2.08 (95%CI 1.29–3.37, p = 0.003), and 0.26 (95%CI 0.11–0.61, p = 0.002) between phenotype 2 (n = 142; low risk) and phenotype 3. For phenotypes classification by random forest, AUCs of phenotypes 1, 2, and 3 were 0.736 ± 0.038, 0.815 ± 0.035, and 0.721 ± 0.03, respectively, slightly better than the decision tree. Then, the classifier effectively identified the phenotypes for new patients in the validation dataset with significant difference on survival curves and hazard ratios. Finally, age and creatinine clearance rate were identified as the top two most important predictors. ML could effectively identify patient prognostic phenotypes, facilitating reasonable management and treatment considering prognostic condition.

Discovering Distinct Phenotypical Clusters in Heart Failure Across the Ejection Fraction Spectrum: a Systematic Review

“…We found 34 eligible clustering studies that were performed between 2012 and 2022, and used varying datatypes, clustering methods, and sample sizes (Table 1 ) [ 10 , 17 – 33 , 34 •, 35 – 38 , 39 •, 40 – 42 , 43 •, 44 – 49 ]. Clustering techniques that were used included hierarchical clustering ( n = 14) [ 10 , 22 , 24 , 25 , 27 , 28 , 31 , 34 •, 36 , 39 •, 40 , 45 , 47 , 49 ], LCA ( n = 10) [ 17 , 21 , 26 , 32 , 33 , 35 , 37 , 43 •, 46 , 48 ], PAM ( n = 5) [ 19 , 29 , 30 , 34 •, 38 ], k-means clustering ( n = 5) [ 23 , 34 •, 41 , 42 , 44 ], and model-based clustering ( n = 3) [ 18 , 20 , 34 •]. Dataset sizes ranged from 103 patients to 318,384 patients.…”

“…Dataset sizes ranged from 103 patients to 318,384 patients. Datatypes varied between registry-based data ( n = 6) [ 19 , 26 , 27 , 42 , 43 •, 46 ], cohort data ( n = 7) [ 20 , 22 , 24 , 28 , 30 , 38 , 41 ], EHR data ( n = 9) [ 10 , 23 , 29 , 31 , 34 •, 44 , 47 – 49 ], and trial data ( n = 12) [ 17 , 18 , 21 , 25 , 32 , 33 , 35 – 37 , 39 •, 40 , 45 ], using varying variable types for the clustering such as clinical variables ( n = 31), echocardiographic variables ( n = 7) [ 10 , 18 , 20 , 22 , 23 , 40 , 49 ], biomarkers ( n = 4) [ 24 , 28 , 38 , 41 ], hemodynamic parameters ( n = 1) [ 23 ], and demographic variables ( n = 1) [ 27 ]. The number of variables used for analysis also varied between 8 and 415, and the number of clusters discovered ranged between 2 to 15.…”

“…Still, most studies exceeded the threshold of 100 participants for each discovered subgroup ( n = 29). Description of missingness ranged from not mentioning missing values at all ( n = 11) [ 18 , 19 , 23 , 25 , 30 , 31 , 35 , 36 , 43 •, 44 , 48 ] to reporting percentage of missing for each variable ( n = 12) [ 10 , 20 – 22 , 24 , 26 , 29 , 37 , 39 •, 40 , 46 , 47 ], but also several studies have only given a very global description of the missingness in the dataset usually limited to which variables passed a specific threshold of missingness ( n = 11) [ 17 , 27 , 28 , 32 , 33 , 34 •, 38 , 41 , 42 , 45 , 49 ]. From the studies that describe handling of missingness they either performed complete case analysis ( n = 9) [ 26 , 32 , 33 , 36 – 38 , 39 •, 43 •, 46 ] or imputation ( n = 13) [ 10 , 17 , 18 , 20 , 22 , 24 , 29 , 34 •, 40 – 42 , 48 , 49 ].…”

“…In this systematic review we examined 34 clustering studies in patients with HF, of which 19 studies were exclusively performed in patients with HFpEF. [ 10 , 17 – 33 , 34 •, 35 – 38 , 39 •, 40 – 42 , 43 •, 44 – 49 ] Methodologies and phenotypes showed major heterogeneity in the study designs, including the types and sizes of the datasets, clustering algorithms, and selected variables. None of the clustering studies fulfilled all components of the quality assessment, however the degree of methodological limitations differed between the studies.…”

Risk of Mortality Prediction Involving Time-Varying Covariates for Patients with Heart Failure Using Deep Learning

“…The data collection was the same as that explained in a previous study [ 14 ]. Medical records of patients with HF hospitalized at Toho University Ohashi Medical Center between April 2016 and March 2019 were reviewed by cardiovascular specialists, constituting a development dataset.…”