Integrating explainable AI with clinical risk factors and ECG classification improved obstruction CAD prediction, with AI-CACS achieving AUC of 0.90 and PPV increasing from 0.37 to 0.61.
Does integrating explainable AI with ECG data and clinical risk factors improve pre-test probability estimation for obstructive CAD compared to clinical variables alone?
Integrating explainable AI with clinical risk factors and ECG classification significantly improves the accuracy of pre-test probability models for detecting obstructive CAD.
Absolute Event Rate: 0% vs 0%
Abstract Background Coronary computed tomography angiography (CCTA) is a non-invasive diagnostic modality for assessing coronary artery disease (CAD), with a growing global trend favouring a "CT-first" approach. However, this strategy strains healthcare systems as demand for CCTA increases. Pre-test probability (PTP) models for obstructive CAD are used to optimize the usage of CCTA. However, these models perform inconsistently across different subpopulations and offer limited discriminatory accuracy, with low positive predictive values (PPV). This study aims to enhance the PTP models with explainable artificial intelligence (XAI) on electrocardiogram (ECG) data. Methods Patients who underwent CCTA for suspected CAD with both CAD-Reporting and Data System (CAD-RADS) classification and Coronary Artery Calcium Score (CACS) were included. Obstructive CAD was defined as CAD-RADS 3 or above. Missing values were imputed. The Risk Factor Weighted Clinical Likelihood (RF-CL) model, based on clinical variables (age, sex, symptom type, smoking, hypertension, dyslipidaemia, diabetes, and family history) was computed. Two gradient descent decision tree models, "AI-ECG" and "AI-CAC," were developed for the pilot study. AI-ECG used clinical risk factors and ECG classification as abnormal, borderline or normal, while AI-CAC further incorporated CACS. We used a cutoff value of 15%, reflecting current guidelines. Comparative analysis was performed using area under the curve (AUC) analysis, the DeLong test for statistical comparison, and feature importance using SHAP values. We intend to develop and present a model based on ECG signal data during the conference. Pilot results We identified 6 178 patients who underwent CCTA between 2022 and 2024 at the Cardiology Centers Netherlands (CCN), of which 2 083 were included (mean age 57.4 ± 10.2 years; 48.1% male). Obstructive CAD was found in 413 (19.8%) patients. The RF-CL model was outperformed by both AI-ECG (AUC: 0.69 vs. 0.80, p0.001) and AI-CACS (AUC: 0.69 vs. 0.90, p0.001). Sensitivity increased in both AI models, from 0.46 to 0.87 for AI-ECG and to 0.83 for AI-CACS, while PPV improved only in the AI-CACS model (from 0.37 to 0.61). Key predictive features were CACS, age, sex, ECG classification, and family history. Conclusion and Discussion Our pilot study shows that integrating XAI with clinical risk factors and ECG classification enhances PTP estimation for obstructive CAD. Implementing this model can better stratify patients at risk of obstructive CAD, support more focused CCTA use and reduce unnecessary testing. Nonetheless, further external validation is needed to assess model performance in real-world, prospective clinical settings. Additionally, models based on raw ECG signals have the potential to extract more detailed information.Reciever Operating Characteristic CurvesFeature Importance
Hennecken et al. (Thu,) reported a other. Integrating explainable AI with clinical risk factors and ECG classification improved obstruction CAD prediction, with AI-CACS achieving AUC of 0.90 and PPV increasing from 0.37 to 0.61.