Artificial intelligence models for predicting cardiovascular diseases in people with type 2 diabetes: A systematic review

People with type 2 diabetes have a higher risk of cardiovascular disease morbidity and mortality. We aim to distil the evidence, summarize the developments, and identify the gaps in relevant research on predicting cardiovascular disease in type 2 diabetes people using AI techniques in the last ten years. A systematic search was carried out for literature published between 1st January 2010 and 30th May 2021 in five medical and scientific databases, including Medline, EMBASE, Global Health (CABI), IEEE Xplore and Web of Science Core Collection. All English language studies describing AI models for predicting cardiovascular diseases in adults with type 2 diabetes were included. The retrieved studies were screened and the data from included studies were extracted by two reviewers. The survey and synthesis of extracted data were conducted based on predefined research questions. IJMEDI checklist was used for quality assessment. From 176 articles identified by the search, 5 studies with sample sizes ranging from 560 to 203,517 met our inclusion criteria. The models predicted the risk of multiple cardiovascular diseases over 5 or 10 years. Ensemble learning, particularly random forest, is the most used algorithm in these models and consistently provided competitive performance. Commonly used features include age, body mass index, blood pressure measurements, and cholesterol measurements. Only one study carried out external validation. The area under the receiver operating characteristic curve for derivation cohorts varied from 0.69 to 0.77. AI models achieved better performance than conventional models in some specific scenarios. AI technologies seem to show promising performance (AUROC in external validation: 0.75 compared to 0.69 from conventional risk scores) for cardiovascular disease prediction in type 2 diabetes people. However, only one of the reviewed models conducted an external validation. Quality of reporting was low in general, and all models lack reproducibility and reusability.