Abstract / Summary
Background/Objectives: Cardisiography is an artificial intelligence-enabled resting vectorcardiography (AI-VCG) system, but evidence for its accuracy in coronary artery disease (CAD) remains limited. We evaluated the AI-VCG-derived CSG-index for obstructive CAD detection, association with stenosis severity, incremental discrimination beyond clinical variables, and concordance with exercise electrocardiography (ECG). Methods: In this two-centre retrospective diagnostic-accuracy study, 721 Cardisiography records from January 2023–December 2025 were screened and 224 patients were included. Invasive coronary angiography (ICA) was the primary reference (n = 172); coronary computed tomography angiography (CCTA) contributed only reference-negative cases (n = 52) to the hierarchical cohort. Obstructive CAD was defined as ≥50% stenosis and CSG-index ≥ −0.27 as positive. Primary analysis assessed obstructive CAD discrimination in the ICA cohort; secondary analyses assessed severe CAD (≥70%), hierarchical-cohort performance, stenosis severity, incremental value beyond available clinical variables and exercise ECG concordance. Results: Obstructive CAD was present in 76/172 patients. In the ICA cohort, CSG-index area under the curve (AUC) was 0.820, with 80.3% sensitivity and 78.1% specificity; AUCs were 0.843 in the hierarchical cohort and 0.785 for severe CAD. CSG-index increased across stenosis classes. After adjustment for available clinical variables, it remained independently associated with obstructive CAD and increased model AUC from 0.806 to 0.886. In the exploratory exercise ECG subgroup (n = 83), agreement was 66.3%, with fewer false-positive CSG-index classifications. Conclusions: The locked CSG-index showed balanced performance for obstructive CAD, provided graded information across stenosis classes, and added discrimination beyond available clinical variables. These findings support prospective multicentre evaluation as a complementary non-invasive resting test.