Artificial Intelligence (AI) Models for Cardiovascular Disease Risk Prediction in Primary and Ambulatory Care: A Scoping Review

BackgroundMortality from cardiovascular disease (CVD) has seen a dramatic increase over the past decades, which has led to a significant increase in the development of risk prediction models. AI-based models have been proposed as a method of enhancing traditional risk models. This review aims to describe the present state of AI risk prediction models for cardiovascular disease in primary and ambulatory care research, and in particular to determine: the stage of development these models have reached, the AI approaches used, and identifying possible sources of bias or limitations in the AI models. MethodsUsing the Arksey and OMalley scoping review method, this review searched Pubmed, EBSCOHost, and Web of Science databases between 2019 and 2024, and relevant studies were identified. Data extraction was performed on eligible included studies. Results22,860 studies were screened, and 25 articles were identified. There was a lack of external validation (20% of models) and lack of clinical impact studies (0%) in this review. A variety of AI techniques were used. Both data and algorithmic biases were commonly identified. There was a lack of geographic variation in datasets (60% were based in the USA) and only 32% of studies reported race and ethnicity data. There was poor predictor and outcome standardization. Calibration was only reported in 24% of models. ConclusionFindings from this review highlight the lack of clinical impact studies and risk of bias in current AI based models. It provides evidence for future refinement and development of AI risk prediction models in cardiovascular disease. Author summaryCardiovascular disease (CVD) is a leading cause of mortality worldwide. In this paper, we undertake a scoping review of artificial intelligence (AI) models for predicting CVD risk in primary care (PC) and ambulatory care (AC) settings. Our findings underscore the critical role of AI in improving the prediction and prevention of cardiovascular diseases, while also highlighting the existing gaps and challenges in the application and validation of these models in primary and ambulatory care. We identified major biases in data from which models were developed, such as geographic discrepancies and inconsistencies among CVD predictors. Additionally, biases in the algorithm were significant issues that need to be addressed in the future. We emphasize the need for multi-centered, standardized, and externally validated models to ensure their utility and reliability in clinical settings. Our research underscores the importance of addressing both data and algorithmic biases, and the necessity of external validation and clinical trials to enhance the accuracy and clinical utility of these AI models in real-world settings.