Incremental predictive value of sex specific risk within pre-test probability models for coronary CT angiography using explainable artificial intelligence
J Hennecken, W R Van de Vijver, G A Somsen, I I Tulevski, R N Planken, B E P M Claessen, M M WinterAbstract
Background
Coronary computed tomography angiography (CCTA) has emerged as a powerful non-invasive imaging modality for the evaluation of coronary artery disease (CAD), with a global trend towards a "CT-first" strategy in stable patients with suspected CAD. Pre-test probability (PTP) models stratify risks of CAD for appropriate utilization of CCTA. However, these models are largely derived from male-dominant cohorts and often underestimate CAD risk in women.
Purpose
This study uses explainable artificial intelligence (AI) for exploration of the incremental predictive power of sex-specific risk (SSR) factors for PTP models.
Methods
Patients undergoing CCTA for suspected CAD at Cardiology Centers Netherlands were included. The ESC RF-CL model was constructed using raw weights and interaction coefficients. The AI-based SSR-PTP model incorporated classic cardiovascular risk factors and additional SSR factors: thyroid/rheumatic disease, migraines (diagnosed ≤20 years), early menopause (≤40 years), and multiple miscarriages. Missing data (<10%) was imputed via nearest neighbors. A gradient boosting decision tree model (70/30 train-test split, depth 5, learning rate 0.01, 5-fold cross-validation) with isotonic regression calibration was developed. Performance was assessed using discriminatory analysis, bootstrapped 95% confidence intervals (CI), net reclassification index (NRI), and explainable SHAP features.
Results
Among 4,234 patients (52% women), 69% had CAD and 21% obstructive disease (mean PTP: 0.21 ± 0.09). Additional SSR data was available for 725 patients. The RF-CL model slightly underperformed in women (AUC 0.71 vs 0.74; mean difference (MD): −0.02, 95%CI [−0.06-0.01]; p = 0.23). In the women’s program subset, the AI SSR-PTP model significantly outperformed RF-CL (AUC 0.89 vs 0.68; MD: 0.21, 95%CI 0.12–0.30; p < 0.01), with improved sensitivity (0.78 vs 0.63) and specificity (0.77 vs 0.63). The model was well calibrated (p = 0.74), and reclassification improved (NRI 0.28, 95% CI 0.11–0.45; p = 0.02). SHAP analysis suggested protective associations for thyroid disease and presence of anemia, and increased risk with rheumatic disease or early-life migraines.
Conclusion
In this large CCTA cohort, integrating SSR factors into an AI-based model markedly enhanced CAD risk prediction and individualized PTP estimation, underscoring the need for sex-sensitive approaches in CAD diagnostics.ROC curve analysisFeature Importance analysis