Heart Failure Risk Prediction With Machine Learning in Atrial Fibrillation

A machine-learning (ML) prediction model has found predictors of heart failure (HF) risk among patients with atrial fibrillation (AF), as well as a way to stratify risk for HF, according to study findings published in the Journal of the American College of Cardiology: Asia.

Researchers sought to create and certify an ML prediction model for HF hospitalization risk in patients with AF. The primary outcome was incidences of hospitalization for HF during the follow-up period, which was until death.

The researchers used the Fushimi AF Registry (a community-based prospective survey of patients with AF in Fushimi-ku, Kyoto, Japan) to divide patients into derivation (N=2383) and validation (N=2011) cohorts. The derivation cohort was used to create an ML model to predict the incidence of hospitalization for HF. The validation cohort was used to test the predictive ability of the model. Follow-up data was obtained for these 4394 patients (mean age, 73.6±10.9 years; 40% women) through April 2019.

Cohorts were well-matched for baseline characteristics of sex, age, body mass index, systolic blood pressure, and body weight. Medical histories were similar between groups for pre-existing HF and history of stroke, valvular heart disease, and coronary artery disease. They differed significantly in history of hypertension, dyslipidemia, and chronic kidney disease.

Overall, 606 patients were hospitalized for HF during a median follow-up of 4.4 years (IQR, 2.1-7.0 years). HF hospitalization annual incidence rate in the derivation cohort was 4.0% per person-year, and in the validation cohort 2.5% per person-year. They observed all-cause death in 22% of all patients, with an annual mortality rate of 6.1% in the derivation cohort and 3.7% in the validation cohort.

The researchers created 6 ML models (random forest, neural network, linear support vector machine, elastic net, light gradient boosting machine, and naïve Bayes) and all models found significant predictive variables in data of transthoracic echocardiography and biomarkers.

No clear criteria for variable selection stood out, so researchers extracted variables based on feasibility, validity, and applicability from the clinician’s perspective and selected 7 variables in the random forest algorithm for the practical ML model.

This random forest algorithm ML model showed a high prediction performance (area under the receiver operating characteristics curve [AUC], 0.75) using variables of cardiothoracic ratio on x-ray, left ventricular [LV] ejection fraction, LV end-systolic diameter, LV asynergy, creatinine clearance, history of HF, and age. This proved significantly superior to the Framingham HF risk model (AUC, 0.67; P <.001).

Researchers noted risk for HF hospitalization during the follow-up period could be stratified with the ML model, based on Kaplan-Meier curves (log-rank; P <.001).

Study limitations include the observational design lacking causative power and the exclusion of potential important variables due to missing data. There is also failure to collect detailed echocardiographic data including at HF hospitalization and lack of external validation.

“ML algorithms had a high predictive performance for HF hospitalization in patients with AF,” the researchers wrote. “Our ML model using 7 simple and readily available variables was able to stratify the risk of hospitalization for HF in patients with AF, providing opportunities for the implementation of strategies to prevent HF among patients with AF.”

Disclosure: The Fushimi AF Registry is supported by Boehringer Ingelheim, Bayer Healthcare, Pfizer, Bristol-Myers Squibb, Astellas Pharma, AstraZeneca, Daiichi Sankyo, Novartis Pharma, MSD, Sanofi-Aventis, and Takeda Pharmaceutical. One study author declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.