Machine Learning for Mortality Prediction in Patients With Heart Failure With Mildly Reduced Ejection Fraction

Figures

• 

  Figure 1. Flowchart for feature selection and model development. a14 benchmark predictors: age, sex, SBP, Na, NYHA III/IV, eGFR*, ischemic heart failure, diabetes, anemia, NT‐proBNP, ACEIs/ARBs treatment, β‐blockers treatment, LVEF, and hemoglobin. ACEI indicates angiotensin‐converting enzyme inhibitor; ARB, angiotensin receptor blocker; AUC, area under the receiver operating characteristic curve; AUPRC, area under the precision‐recall curve; BMI, body mass index; BSA, body surface area; CBCV, Cox Boost with 10‐fold cross‐validation; eGFR, estimated glomerular filtration rate; HFmrEF, heart failure with mildly reduced ejection fraction; L, lymphocyte; LVEF, left ventricular ejection fraction; N, neutrophil; NT‐proBNP, N‐terminal probrain natriuretic peptide; NYHA, New York Heart Association; RF, random forest; SBP, systolic blood pressure; SVM, support vector machine; and XGBoost, eXtreme Gradient Boosting. *eGFR was calculated from the creatinine by the chronic kidney disease epidemiology collaboration (CKD‐EPI) equation.

• 

  Figure 2. Model prediction performance and clinical usefulness of the “All‐in” MLBPMs. “All‐in” MLBPMs showed good discrimination performance by ROC (A) and PR curves (B) and higher clinical net benefit (C). AUC indicates area under the receiver operating characteristic curve; AUPRC, area under the precision‐recall curve; MLBPM, machine‐learning‐based prediction model; PR, precision‐recall; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; and XGBoost, eXtreme Gradient Boosting.

• 

  Figure 3. Prediction distribution of “All‐in” MLBPMs in patients with HFmrEF. A, The XGBoost model. B, The RF model. C, The SVM model. D, The reference logistic model. The machine learning model could be more precise in stratifying patients at risk of all‐cause mortality by positively clustering patients, in order of risk, who died during long‐term follow‐up. The prediction distributions were consistent between the whole HFmrEF population and those with LVEF 41% to 49%. HFmrEF indicates heart failure with mildly reduced ejection fraction; LVEF, left ventricular ejection fraction; MLBPM, machine‐learning‐based prediction model; RF, random forest; SVM, support vector machine; and XGBoost, eXtreme Gradient Boosting.

• 

  Figure 4. Model prediction performance and clinical usefulness of the “CBCV” MLBPMs. “CBCV” MLBPMs showed good discrimination performance by ROC (A) and PR curves (B) and higher clinical net benefit (C). AUC indicates area under the receiver operating characteristic curve; AUPRC, area under the precision‐recall curve; CBCV, CoxBoost with 10‐fold cross‐validation; MLBPMs, machine‐learning‐based prediction models; ROC, receiver operating characteristic; PR, precision‐recall; RF, random forest; SVM, support vector machine; and XGBoost, eXtreme Gradient Boosting.

• 

  Figure 5. Feature importance and SHAP analysis of the “All‐in” XGBoost model. Feature importance plot of the “all‐in” XGBoost (A) and the explanation by SHAP analysis (B). SHAP analysis was helpful in understanding the direction of a specific feature to a higher or lower prediction, and the color indicates whether that feature is high (in purple) or low (in yellow) for that observation. ACEI indicates angiotensin‐converting enzyme inhibitor; ALB, albumin; ALT, alanine transaminase; ALP, alkaline phosphatase; Apo (A), apolipoprotein; APTT, activated partial thromboplastin time; ARB, angiotensin receptor blocker; AST, aspartate aminotransferase; BMI, body mass index; BUN, blood urea nitrogen; CK, creatine kinase; Cr, creatinine; DBP, diastolic blood pressure; DM, diabetes mellitus; ESR, erythrocyte sedimentation rate; FIB,fibrinogen; FT3, free triiodothyronine; FT4, free thyroxine; GGT, γ‐glutamyl transpeptidase; Hb, hemoglobin; HDL, high‐density lipoprotein; HR, heart rate; Hs‐CRP, high‐sensitivity C‐reactive protein; IHF, ischemic heart failure; IVS, interventricular septum; LADD, left atrial diastolic diameter; LDH, lactate dehydrogenase; LVEDD, left ventricular end‐diastolic diameter;…