1290-P: Developing Machine Learning Model for Predicting Acute Coronary Syndrome in Type 2 Diabetes Mellitus Patients through Substitution of Propensity Scores for Binary Variables

Introduction & Objective: T2DM patients have a high risk for acute coronary syndrome (ACS), requiring early and accurate risk stratification. However, traditional risk models lack generalizability for this patient population, while machine learning (ML)-based models have been plagued by inaccuracies in the training data, particularly those of binary information. Our study aims to develop an ML model for ACS prediction in T2DM patients by applying a novel data representation approach.

Methods: We used the EMRs of T2DM patients treated with basal insulin. The study unit was treatment period, during which the subjects were continuously treated with basal insulin. The primary outcome was ACS occurrence during this period. We addressed EMR inaccuracies by replacing binary information with their respective propensity scores (PS), re-defined as the probability of having a record given covariates. We used various ML algorithms to develop ACS prediction models and evaluated their performance. Additionally, the Shapley Additive Explanation (SHAP) method identified important clinical predictors of ACS.

Results: The study included 9,338 patients (mean age 60.2 years and 56.6% male) over 10,184 treatment periods. The most prevalent comorbidities were hypertension and dyslipidemia. Notably, 6.9% experienced ACS. Our best performing model achieved a 0.969 in all performance metrics. Compared with models developed via traditional methodologies, our approach improved prediction performance by over 6 percentage points. SHAP analysis showed that older age, higher baseline glucose, history of antithrombotic therapy, history of chest pain, and indicators of T2DM progression (e.g., senile cataract) were important ACS risk factors.

Conclusion: Our model showed a high generalizability for T2DM patients, improved prediction accuracy, and higher reliance on clinical predictors that align with current medical understanding.