Prognostic Implications of Coronary CT Angiography: 12-Year Follow-Up of 6892 Patients

OBJECTIVE. The purpose of this study was to add to evidence of the long-term prognostic value of coronary CT angiography (CCTA) when combined with nonimaging risk factors and to investigate how CCTA can contribute to the decision to start statin therapy. MATERIALS AND METHODS. Patients underwent CCTA in an outpatient setting for a variety of indications. The National Death Index from February 2004 through December 2018 was queried to identify the outcomes of death due to any cause (all-cause mortality) and death due to coronary artery disease. Framingham and machine learning risk estimation models were constructed. Machine learning inputs were generated from radiologists' descriptions of the findings on structured report forms and not directly from the image pixels. Kaplan-Meier survival curves and Cox proportional hazards were calculated. Clinical benefit was assessed on the basis of the potential impact on assignment of statin therapy. RESULTS. A total of 6892 outpatients were studied, 4452 (64.6%) of whom were men (mean [± SD] age, 51.2 ± 11.1 years) and 2440 (35.4%) of whom were women (mean age, 57.3 ± 12.2 years). The median follow-up was 11.9 years. Among the 6892 patients, 569 deaths (8.3%) were attributed to all-cause mortality, and 94 deaths (1.4%) were due to coronary artery disease. Survival showed strong dependence on the extent of coronary atherosclerosis. For all-cause mortality, the AUC was 0.85 (95% CI, 0.83-0.86) for the machine learning risk estimation model versus 0.79 (95% CI, 0.78-0.81) for the Framingham risk estimation model (p < 0.001), and for death due to coronary artery disease, the AUC was 0.87 (95% CI, 0.84-0.91) for the machine learning model versus 0.82 (95% CI, 0.77-0.86) for the Framingham model (p = 0.004). Using machine learning risk estimates, the prescription of statins could more accurately be matched to the burden of coronary disease than when Framingham risk estimates were used. CONCLUSION. Compared with the Framingham model, the machine learning model improved risk estimation. Similar models might be useful to better target prescription of statins and reduce their overuse.