Using AI-Quantitative CT to evaluate the relationship between coronary artery calcium and segment involvement scores in quantifying coronary plaque burden.

PURPOSE: Accurate assessment of coronary plaque burden is essential for risk stratification in coronary artery disease (CAD). The Coronary Artery Disease – Reporting and Data System (CAD-RADS) 2.0 classification incorporates P-scores derived from coronary artery calcium (CAC) and segment involvement scores (SIS) to semi-quantitatively characterize plaque burden. However, limited data exist on the concordance and clinical implications of these two plaque characteristics.

METHODS

We retrospectively analyzed 461 coronary CT angiography (CCTA) studies using a commercial AI-based quantitative CT (AI-QCT) plaque analysis platform (CLEERLY). CAC scores were extracted using prompt-based natural language processing, and SIS were determined by AI quantification. Unilateral P-scores for CAC (PCAC) and SIS (PSIS) were calculated according to CAD-RADS 2.0 thresholds. Agreement between PCAC and PSIS was assessed using Cohen’s kappa, Wilcoxon signed-rank test, and McNemar’s test. Subgroup analyses evaluated distribution of P-score categories, sex and age in CAC = 0 patients, and the impact of symptom status using Wilcoxon rank-sum, chi-square, Fisher’s exact test, and ordinal logistic regression.

RESULTS

The median patient age was 67 years, and 38.7% were female. CAC and SIS produced concordant P-scores in 23% of cases, with discordance in 77%. SIS determined the final P-score in 75.5% and CAC in only 1.5%. Cohen’s kappa indicated modest agreement (κ = 0.40, p < 0.01), while Wilcoxon signed-rank and McNemar’s tests revealed significant discordance, with SIS tending to assign higher P-score categories (p < 0.01). Among patients with CAC = 0 (n = 97; median age 62; 65% female), 95% had non-calcified plaque by AI-QCT, and only 4 had both CAC and SIS of zero. P-score, PCAC, and PSIS distributions did not differ significantly by symptom status (all p > 0.05), although a significant difference was detected for PSIS on chi-square and Fisher’s exact tests but not by ordinal regression or Wilcoxon rank-sum. Further analysis demonstrated that discordance was most prominent in higher P-score categories and was primarily driven by non-calcified plaque.

CONCLUSION

Our study demonstrates frequent discordance between CAC and SIS in CAD-RADS 2.0 plaque burden classification, particularly due to the detection of non-calcified plaque by SIS. Reliance on either CAC or SIS alone may result in risk misclassification, especially for patients with CAC = 0, the majority of whom had non-calcified plaque by AI-QCT. These findings highlight the need for standardized criteria and integration of both measures in automated plaque quantification to improve cardiovascular risk assessment, especially in discordant populations.