TabM para diagnóstico de doença cardíaca: ensembling eficiente de redes neurais em dados tabulares clínicos

Abstract

Cardiovascular diseases represent the main cause of global mortality, being responsible for approximately 19.8 million deaths annually, making early diagnosis fundamental to reducing mortality and improving patients' quality of life. This work investigated the application of the TabM model, a modern deep learning architecture based on dense neural networks (MLPs) and parameter-efficient ensembling that, despite not being the absolute state of the art, demonstrated unique resilience to overfitting among tabular deep learning models and achieved 2nd place ranking in the TabArena benchmark, for diagnosing heart disease using the Cleveland Heart Disease dataset. The objective was to evaluate whether this competitive architecture in general tabular benchmarks would achieve competitive performance with the best results reported in the literature for cardiac diagnosis. The methodology involved selection of relevant clinical features, application of class balancing techniques, hyperparameter optimization, and stratified cross-validation. The results did not fully support the experimental hypotheses, leading to the acceptance of the null hypothesis. The model demonstrated robustness in AUC-ROC, however it fell below the established criteria in accuracy, recall, and F1-score. Beyond the experimental validation, the study delivered a functional Minimum Viable Product (MVP) in Streamlit for interactive clinical inference, a significant differentiator compared to related works that did not provide operational tools. This work represents the first known application of TabM to heart disease diagnosis, establishing a benchmark for future research and a reproducible open-source pipeline.