Imaging- and ECG-based prediction tools demonstrated excellent discrimination for valvular heart disease (C-statistics 0.858 and 0.821, respectively), but lack external validation.
Meta-Analysis
Do prediction models accurately identify individuals with underlying valvular heart disease?
Imaging and ECG-based prediction tools demonstrate excellent discrimination for detecting valvular heart disease, though prospective validation is needed before clinical implementation.
Effect estimate: C-statistic 0.858 (imaging models) (95% CI 0.821-0.905)
Abstract Background Valvular heart disease (VHD) is a leading cause of cardiovascular morbidity and mortality, yet many high-risk individuals remain undiagnosed until symptomatic. The 2025 European Society of Cardiology guidelines lack structured recommendations for early detection or risk-based screening, though intervention in asymptomatic patients is suggested. Prediction models - including imaging, electrocardiogram (ECG), multivariable, and biomarker-based approaches - could help identify high-risk individuals for timely diagnosis and intervention. Purpose To provide an overview of prediction tools for VHD by model type and valve disease. Methods We conducted a systematic review of adult VHD prediction tools -including imaging, ECG, clinical, and biomarker-based models - using MEDLINE and Embase through 5 November 2024. Models with C-statistic/AUROC data from ≥3 cohorts were pooled via Bayesian meta-analysis, stratified by model type or valve class, with heterogeneity assessed using 95% prediction intervals. Discrimination was classified a priori as inadequate (0.60), adequate (0.60-0.70), good (0.70-0.80), or excellent (0.80). Results From 4,667 records, 17 studies (23 cohorts) reporting 29 prediction tools were included: 10 imaging-based, 10 ECG-based, 3 multivariable clinical, 3 biomarker-based, and 1 genetic-based. Twenty-four used advanced computational (machine learning n=5; deep learning n=19). Imaging- and ECG-based models showed excellent discrimination (0.858 95% CI 0.821-0.905 and 0.821 95% CI 0.786-0.853), while clinical and biomarker models were acceptable (0.788 95% CI 0.753-0.881 and 0.704 95% CI 0.550-0.858)(Figure 1). At the lesion level, MS and MR models demonstrated excellent discrimination (0.855 95% CI 0.718-0.968 and 0.886 95% CI 0.824-0.929), while AS and AR models showed excellent and acceptable performance (0.814 95% CI 0.762-0.859 and 0.785 95% CI 0.785 (0.656-0.862)(Figure 2). Meta-analysis of individual tools was not possible due to insufficient external validation, and clinical utility or prospective validation was lacking. Conclusion Imaging- and ECG-based prediction tools show strong potential for predicting underlying VHD. However, limited external validation and clinical impact data restrict translation to clinical practice. Prospective evaluation in risk-based screening programs are needed to enable timely diagnosis and intervention.Figure 1For image description, please refer to the figure legend and surrounding text. Figure 2For image description, please refer to the figure legend and surrounding text.
Helbitz et al. (Mon,) conducted a meta-analysis in Valvular heart disease. Prediction models for valvular heart disease was evaluated on Model discrimination (C-statistic/AUROC) (C-statistic 0.858 (imaging models), 95% CI 0.821-0.905). Imaging- and ECG-based prediction tools demonstrated excellent discrimination for valvular heart disease (C-statistics 0.858 and 0.821, respectively), but lack external validation.