Skip to main content
. 2022 Jan 26;10(2):232. doi: 10.3390/healthcare10020232

Table 3.

Summary of literature regarding mortality outcomes using CCTA data.

Study Study Design/Sample Size ML Brief Description and Outcomes Results Limitations
Motwani et al. [195]
2016		 Multicenter prospective study, 10,030 patients with suspected CAD Boosted ensemble (LogitBoost) 25 clinical and 44 CCTA parameter used to create ML model
Outcome: Prediction of 5-year ACM; compared against clinical risk scores and CCTA parameters. 		AUC: ML (0.79) > Segment stenosis score (SSS) (0.64) and FRS (0.61); p < 0.001.

  1. Observational; concern for selection bias

  2. Cardiac-specific endpoints were not defined, given the data unavailability.
Hoshino et al. [198]
2016		 Multicenter retrospective study, 220 patients with intermediate LAD stenosis Unsupervised hierarchical clustering Two clusters (CS1 and CS2) using 42 variables created via ML.
Outcome:

  1. Relation between FAI and CCTA defined clusters,

  2. Prognostic value of ML-derived clusters in combination with FAI.

  1. Age, CS1 features (higher plaque volume, remodeling index, higher FAI amongst others), and FAI were independent predictors of MACE.

  2. Improved NRI with (FRS + CS1 + FAI) as compared to FRS alone.

  1. Retrospective, small size

  2. Majority of vessels were LAD; hence the study was restricted to a specific population.

  3. 40% cardiac events were non-LAD revascularization; hence the results were not generalizable.
Van Rosendael et al. [197]
2018		 Multicenter prospective study, 8844 patients with suspected CAD Boosted ensemble 35 variables (SS and plaque composition for 16 coronary segments and 3 additional variables) compared with traditional CT scores.
Outcome: ML vs. traditional CT scores in predicting 5-year composite MI and death. 		AUC for ML (0.77) > SSS (0.70)

  1. No comparison with clinical risk scores

  2. Retrospective study with risk of selection bias
Johnson et al. [194]
2019		 Single-center retrospective study, 6892 patients K nearest neighbors ML model (64 vessel-related features) vs. CAD-RADS.
Outcome: Prediction of ACM, CAD-related deaths. Also, decision to start statin.

  1. AUC for all-cause mortality (0.77) > CAD-RADS (0.72); AUC for CAD-related deaths—ML (0.85) > CAD-RADS (0.79).

  2. Significant increase in sensitivity with ML model.

  1. Retrospective study with limited population diversity

  2. Unblinded CCTA results that might have affected event incidence
Johnson et al. [199]
2020		 Single-center retrospective study, 6892 patients ML model developed via radiologist report.
Outcome: Prediction of ACM and CAD-related mortality; compared against FRS. Also, decision to start statin.

  1. ACM: AUC for ML (0.85) > FRS (0.79) CAD related deaths: AUC for ML (0.87) > FRS (0.82)

  2. Using ML, equally high sensitivity but significant reduction in unnecessary statin prescription (AUC for ML 0.89 vs. FRS 0.75).

  1. Retrospective study design

  2. Concern for misclassification bias due to incomplete follow-up
Tesche et al. [196]
2021		 Single-center retrospective study, 361 patients with suspected and confirmed CAD Boosted ensemble (RUSBoost) 28 clinical, CCTA scores and adverse plaque characteristics included.
Outcome: 5-year MACE prediction; compared against FRS, CCTA scores and adverse plaque features.

  1. AUC for ML (0.96) > AS (0.84) > FRS (0.76).

  2. Important imaging parameters: SSS, obstructive CAD of RCA.

  3. Important clinical factors: age, FRS

  1. Small sample size, retrospective study design

  2. Follow-up using medical records

  3. No external validation to test prognostic accuracy

ACM: all-cause mortality; AS: Agatston score; CAD-RADS: coronary artery disease reporting and data system; CS: cluster sample; FAI: fat attenuation index; FRS: Framingham risk score; RCA: right coronary artery; SSS: segment stenosis score.