JMIR Publications

ABSTRACT

Background:

Transcatheter aortic valve replacement (TAVR) is a commonly used treatment for severe aortic stenosis. As degenerative aortic stenosis is primarily a disease of older adults, frailty assessment is essential to patient selection and optimal periprocedural outcomes

Objective:

This study aimed to enhance frailty assessment by integrating real-world structured and unstructured data for automated frailty assessment in transcatheter aortic valve replacement (TAVR) candidates

Methods:

This study analyzed data from 14,000 patients to assess frailty in TAVR patients between January 2018 and December 2019 to assess frailty in TAVR patients at the University of Florida. Frailty was identified using the Fried criterion, including weight loss, exhaustion, walking speed, grip strength, and physical activity. Latent Dirichlet Allocation (LDA) for topic modeling and XGBoost for frailty prediction were applied to unstructured clinical notes and structured electronic health records (EHR) data. We also used LASSO regression for feature selection. Model performance was rigorously evaluated using nested cross-validation, ensuring the generalizability of the findings

Results:

The model’s performance was significantly improved by combining unstructured clinical notes with structured EHR data, achieving an area under the curve (AUC) of 0.82, which surpassed the EHR-only model’s AUC of 0.64. The SHAP analysis found that congestive heart failure management, back problems, and atrial fibrillation were the top frailty predictors. Additionally, the LDA topic modeling identified seven key topics, highlighting the role of specific medical treatments in predicting frailty

Conclusions:

Integrating unstructured clinical notes and structured EHR data led to a notable enhancement in predicting frailty. This method shows great potential for standardizing frailty assessment using RWD and improving patient selection for TAVR