Background: Coronary artery calcification (CAC) is a critical marker of subclinical atherosclerosis, yet its primary diagnostic tool, coronary computed tomography (CT), is limited by cost and radiation. We aimed to develop and validate a robust machine learning framework to identify significant CAC (Agatston score ≥100) using features extracted from routine 12-lead electrocardiograms (ECGs), including legacy archives stored as PDFs. Methods: This retrospective study utilized 382 consecutive patient records. A novel signal-processing pipeline was developed to convert PDF-based ECGs into standardized 500-Hz digital waveforms (WFDB format). From these signals, 360 morphological and temporal features were extracted using the NeuroKit2 library. To ensure clinical reliability and prevent optimistic bias, a Balanced Random Forest (BRF) classifier was integrated into a repeated “inner-outer” nested cross-validation framework combined with Minimum Redundancy Maximum Relevance (mRMR) feature selection. Results: The mRMR-BRF model demonstrated high diagnostic accuracy with a grand mean AUC of 0.8211 (95% CI: 0.8018–0.8404), outperforming standard algorithms. Explainability analysis via SHAP identified age, ST-segment level in lead V4, T/QRS ratio in lead III, and QRS duration dispersion as the most significant predictors. These features suggest that latent electrical signatures of subclinical myocardial remodeling and impaired perfusion are detectable via AI-enhanced ECG analysis. Conclusion: Our framework establishes digitized 12-lead ECGs as a scalable and non-invasive digital biomarker for significant CAC. By enabling the quantitative analysis of legacy PDF archives and providing high predictive stability through nested validation, this approach offers a cost-effective pre-screening tool to prioritize high-risk patients for definitive imaging and early preventive intervention.