Machine learning algorithms identifying the risk of new-onset ACS in patients with type 2 diabetes mellitus: A retrospective cohort study

Abstract

Background: In recent years, the prevalence of type 2 diabetes mellitus (T2DM) has increased annually. The major complication of T2DM is cardiovascular disease (CVD). CVD is the main cause of death in T2DM patients, particularly those with comorbid acute coronary syndrome (ACS). Although risk prediction models using multivariate logistic regression are available to assess the probability of new-onset ACS development in T2DM patients, none have been established using machine learning (ML).

Methods: Between January 2019 and January 2020, we enrolled 521 T2DM patients with new-onset ACS or no ACS from our institution's medical information recording system and divided them into a training dataset and a testing dataset. Seven ML algorithms were used to establish models to assess the probability of ACS coupled with 5-cross validation.

Results: We established a nomogram to assess the probability of newly diagnosed ACS in T2DM patients with an area under the curve (AUC) of 0.80 in the testing dataset and identified some key features: family history of CVD, history of smoking and drinking, aspartate aminotransferase level, age, neutrophil count, and Killip grade, which accelerated the development of ACS in patients with T2DM. The AUC values of the seven ML models were 0.70-0.96, and random forest model had the best performance (accuracy, 0.89; AUC, 0.96; recall, 0.83; precision, 0.91; F1 score, 0.87).

Conclusion: ML algorithms, especially random forest model (AUC, 0.961), had higher performance than conventional logistic regression (AUC, 0.801) for assessing new-onset ACS probability in T2DM patients with excellent clinical and diagnostic value.

Keywords: acute coronary syndrome; machine learning; nomogram; random forest; type 2 diabetes mellitus.