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

Table 2.

Studies evaluating the impact of coronary artery calcium score (CACS) among other variables in the prediction of mortality in patients with no history of coronary artery disease.

Study Study Design/Sample Size ML Model Brief Description and Follow-Up Results Limitations
Eisenberg et al. [187]
2020		 Prospective single-center study,
2068 asymptomatic patients		 Convolutional neural network To check for impact of EAT volume and EAT attenuation computed via deep learning in prediction of MACE, defined as defined as MI, late (>180 days) revascularization and cardiac death.
Follow up: >14 years		 Increased EAT volume (HR: 1.35) and decreased EAT attenuation (HR 0.83) independently associated with MACE in addition to CACS (HR 1.25) and ASCVD score (HR 1.03), p < 0.01 for all.

  1. Study done on asymptomatic patients; external validation needed if applied on symptomatic patients.

  2. Previous-generation CT scanners used (data collected from 1998–2005).
Han et al. [188]
2020		 Retrospective multicenter study,
86,155 asymptomatic patients		 Boosted ensemble ML model with 35 clinical, 32 lab, and 3 CACS parameters (CACS, calcium volume, and calcium mass) in prediction of all-cause mortality
Median follow up: 4.6 years		 ML (0.82) > ASCVD score + CACS (0.74) > Framingham risk score + CACS (0.70)—reported as AUC in the test set.
No statistical difference in the performance in the validation set.

  1. Retrospective

  2. All-cause mortality reported rather than specific cardiac endpoints.
Nakanishi et al. [190]
2021		 Multicenter observational study,
66,636 asymptomatic patients		 Boosted ensemble (Logitboost) ML model incorporating 46 clinical and 31 CT variables—CAC score, extra coronary scores (not including EAT) in prediction of cardiovascular (CHD + stroke + CHF + other circulatory diseases), and coronary heart disease (CHD) deaths
Follow up: 10 years

  1. For cardiovascular deaths: AUC for ML (all) 0.845 > ASCVD (0.821) > CAC score (0.78).

  2. For coronary heart disease deaths: AUC for ML (all) 0.860 > ASCVD (0.835) > CAC score (0.816).

  1. Multiple CT variables, including EAT, were not available for some patients.
Commandeur et al. [186]
2020		 Prospective single-center study,
1912 asymptomatic patients		 Boosted ensemble (XgBoost) ML model using clinical variables, plasma lipid panel measurements, CAC, aortic calcium, and automated EAT measures in prediction of MI and cardiac deaths.
Median follow up: 14.5 years

  1. ML model 0.82 > ASCVD risk score 0.77 ~ CAC 0.77.

  2. Age, ASCVD risk score, and CACS were the three most important features seen in the model.

  1. Overfitting; since small number of events (<4%).

  2. Study done on asymptomatic patients; external validation needed if applied on symptomatic patients.
Tamarappoo et al. [189]
2021		 Prospective single-center study,
1069 asymptomatic patients		 Boosted ensemble (XgBoost) ML model using 12 variables from ASCVD score, 5 CT parameters (including EAT volume and attenuation) and top 15 serum biomarkers) to predict cardiac events
Mean follow up: 14.5 years		 ML (0.81) > CAC (0.75) > ASCVD (0.74).

  1. Single-center study

  2. Overfitting; given the small number of cardiac events during follow up (~2%)

ASCVD: atherosclerotic cardiovascular disease; CHF: congestive heart failure; EAT: epicardial adipose tissue; HR: hazard ratio; MI: myocardial infarction.