Coronary heart disease is the second leading cause of death worldwide. As a preventable and treatable chronic disease, early screening is of great importance for disease control. However, previous screening tools relied on physician assistance, thus cannot be used on a large scale. Many facial features have been reported to be associated with coronary heart disease and may be useful for screening. However, these facial features have limitations such as fewer types, irregular definitions and poor repeatability of manual judgment, so they can not be routinely applied in clinical practice. With the development of artificial intelligence, it is possible to integrate facial features to predict diseases. A recent study published in the European Heart Journal showed that coronary heart disease can be predicted using artificial intelligence based on facial photos. Although this work still has some limitations, this novel technology will be promise for improving disease screening and diagnosis in the future.
ObjectiveTo systematically review the models for predicting coronary artery disease (CAD) and demonstrate their predictive efficacy. MethodsPubMed, EMbase and China National Knowledge Internet were searched comprehensively by computer. We included studies which were designed to develop and validate predictive models of CAD. The studies published from inception to September 30, 2020 were searched. Two reviewers independently evaluated the studies according to the inclusion and exclusion criteria and extracted the baseline characteristics and metrics of model performance.ResultsA total of 30 studies were identified, and 19 diagnostic predictive models were for CAD. Seventeen models had external validation group with area under curve (AUC)>0.7. The AUC for the external validation of the traditional models, including Diamond-Forrester model, updated Diamond-Forrester model, Duke Clinical Score, CAD consortium clinical score, ranged from 0.49 to 0.87.ConclusionMost models have modest discriminative ability. The predictive efficacy of traditional models varies greatly among different populations.