0832 Evaluating Supervised Machine Learning Models for Cardiovascular Disease Prediction Using Conventional Risk Factors, Apnea-Hypopnea Index and Epworth Sleepiness Scale

Abstract

Introduction: Cardiovascular (CV) disease is the leading cause of death and there is a well-established relationship with obstructive sleep apnea (OSA). Accurate CV risk prediction using epidemiological data is fundamental to advance precision medicine. Machine learning methods can improve CV risk prediction, especially when OSA-related risk factors are incorporated. We evaluated the performance of different supervised machine learning methods to predict incident CV disease based on conventional risk factors and OSA severity metrics. Methods: We used data from 3,674 individuals without baseline CV disease from the Sleep Heart Health Study (SHHS) and evaluated the following predictors at baseline: age, sex, race, ethnicity, diabetes, hypertension, HDL, total cholesterol, triglycerides, lipid-lowering medication, alcohol, smoking, apnea-hypopnea index (AHI) and Epworth Sleepiness Scale (ESS). To verify the ability of different machine learning methods to predict incident CV disease (median observation period 11.4 years), we compared naïve Bayes (NB), logistic regression (LR), elastic-net regularized general linear model (enGLM), multi-layer perceptron neural network, decision tree, random forest, extreme gradient boosting, K-nearest neighbors, and support vector machine (SVM). Data was split into training (N=2,939) and testing (N=735). Training data was used to optimize hyper-parameters and evaluate model performance using 5-fold cross-validation. Training and testing performances were measured by the area under the receiver operating characteristics curve (AUC). Results: The average cross-validation training AUC varied between 0.498 (for SVM) and 0.767 (for enGLM). All methods, except SVM, had a significant improvement in disease prediction compared to the null model (Benjamini-Hochberg adjusted p<0.021). The models with the highest performance in the test data were LR (AUC [95%CI]=0.752 [0.712-0.793]), enGLM (0.749 [0.708-0.791]) and NB (0.719 [0.673-0.766]). Age was the most important predictor across all evaluated models. Conclusion: We have demonstrated the applicability of supervised machine learning to predict CV disease in the SHHS using conventional risk factors, AHI and ESS. These findings can be used as benchmark results for comparing future CV analysis over different methods and with other available features.