Full feature selection for estimating KAP radiation dose in coronary angiographies and percutaneous coronary interventions

TL;DR

This study identifies key demographic and clinical features to accurately estimate radiation dose in coronary angiographies and interventions, enhancing clinical dose assessment models.

Contribution

It introduces a feature selection approach to improve KAP radiation dose estimation using support vector regression in interventional cardiology.

Findings

Top features include FN1AC, FN2BA, and patient weight.

Model performance improves with up to 30 features.

Selected features can serve as predictors or in combined models.

Abstract

In interventional cardiology (IC) the radiation dose variation is very significant, and its estimation has been difficult due to the complexity of the treatments. In order to tackle this problem, the aim of this study was to identify the most important demographic and clinical features to estimate Kerma-Area Product (KAP) radiation dose in coronary angiographies (CA) and percutaneous coronary interventions (PCI). The study was retrospective using clinical patient data from 838 CA and PCI procedures. A total of 59 features were extracted from the patient data and 9 different filter-based feature selection methods were used to select the most informative features in terms of the KAP radiation dose from the treatments. The selected features were then used in a support vector regression (SVR) model to evaluate their performance in estimating the radiation dose. The ten highest-ranking features were: 1) FN1AC (CA), 2) FN2BA (PCI), 3) weight, 4) post-stenosis 0%, 5) multi-vessel disease, 6) number of procedures 3, 7) pre-stenosis 100%, 8) American Heart Association (AHA) score C, 9) pre-stenosis 85% and 10) gender. The performance of the SVR model increased (mean squared error ~ 450) with the number of features approximately up to 30 features. The identification of the most informative features for CA and PCI KAP is an important step in determining suitable complexity models for clinical practice. The highest-ranking features can be used as individual predictors of IC procedure KAP or can be incorporated into combined complexity score or different estimation models in the future.