JMIR Publications

ABSTRACT

Background:

Currently, there is a paucity of literature addressing personalized risk stratification using multimodal data in patients with symptomatic aortic stenosis (AS) and heart failure with preserved ejection fraction (HFpEF) following transcatheter aortic valve replacement (TAVR).

Objective:

To enhance the performance of risk assessment models in this patient population, we aimed to develop a predictive model for adverse outcomes using various machine learning (ML) techniques.

Methods:

This multicenter cohort study included 326 patients diagnosed with severe AS and HFpEF who underwent TAVR between January 2017 and December 2023. Patients were allocated to training (n=195) and independent validation (n=131) sets based on hospital affiliation. A dual-phase feature selection process, combining LASSO logistic regression and the Boruta algorithm, was employed to identify relevant variables from the multimodal dataset. Five ML models—decision tree, K-nearest neighbors, random forest, support vector machine, and extreme gradient boosting—were utilized to construct a visualization and explainable predictive framework to elucidate model decision-making processes.

Results:

The primary features identified included age, N-terminal pro-brain natriuretic peptide (NT-proBNP), fasting blood glucose, triglyceride/high-density lipoprotein cholesterol (TG/HDL-C) ratio, triglyceride glucose (TyG) index, TyG-Body Mass Index (BMI) Index, atherogenic index of plasma (AIP) index, and Apolipoprotein B. Among the five models, the support vector machine demonstrated the best predictive performance for major adverse cardiovascular and cerebrovascular events (MACCEs) in patients with severe AS and HFpEF following TAVR, achieving an area under the curve (AUC) of 0.783 (95% confidence interval: 0.666-0.899) in the independent validation set. The model exhibited good calibration and robust predictive power in both training and validation sets and demonstrated the highest net benefit in decision curve analysis compared to other models. To extract significant variables influencing the algorithm and ensure model appropriateness, we interpreted cohort and personalized model predictions using SHAP values.

Conclusions:

Our ML-based multimodal model, incorporating eight readily accessible predictors, demonstrated robust predictive capability for incident MACCEs within one year. This model can be utilized to identify high-risk individuals with AS and HFpEF following TAVR, potentially aiding in risk stratification and personalized treatment strategies. Clinical Trial: The study has been registered with the China Clinical Trials Register (ChiCTR2300075597).