Text-Based Classification of Mitral Valve Disease Severity from Echocardiography Reports

Echocardiography reports frequently describe mitral valve disease severity in free text, which limits secondary data use.

This study aimed to evaluate the feasibility and performance of automated mitral valve disease severity classification from unstructured echocardiography reports.

A total of 6,838 free-text echocardiography reports were processed using classical TF-IDF-based machine learning models and a fine-tuned BioBERT-based transformer model.

TF-IDF-based classification models provided reasonable baseline performance, with the most effective support vector machine achieving an accuracy of 0.85. The BioBERT model demonstrated substantially improved classification performance compared to classical approaches, misclassifying only 13 out of 1,026 test cases and achieving consistently high class-wise accuracies above 0.99 across all severity categories.

Automated classification of mitral valve disease severity from free-text echocardiography reports is feasible. Context-aware language models such as BioBERT substantially improve classification performance and support reliable classification of unstructured clinical text.