Abstract 4143475: Early Dementia Prediction Using Electrocardiogram: Insights from Variational Autoencoder-Derived Features

Introduction: Dementia, characterized by cognitive decline and impaired judgment, imposes a significant economic burden due to its rising prevalence and high diagnostic costs. Recent research has linked dementia to various biomarkers, including cardiac biomarkers like heart rate variability. This study investigates the use of Variational Autoencoder (VAE)-derived features from electrocardiogram (ECG) data for predicting dementia, offering a cost-effective alternative to traditional diagnostics. Hypothesis: Features derived from VAE applied to ECG data can be utilized to predict dementia. Methods: ECG data from Sinchon and Yongin Severance Hospital, South Korea, were used to develop a VAE-based predictive model. Patients diagnosed with dementia and matched controls without dementia were included, using propensity score matching for age and gender. For the dementia-develop group, data from the 5 years preceding the first diagnosis were used; for the dementia-free group, the ECG date at the median age served as the index date. Each median ECG was processed through the VAE to derive 32 features. A Deep Neural Network (DNN) was used to predict dementia 5 years before the index date, with Sinchon data for training and Yongin data for testing. Trajectory analysis of VAE features revealed distinct trends between the groups. Results: The model achieved AUROCs of 0.749 (training) and 0.701 (testing), and AUPRCs of 0.490 (training) and 0.358 (testing). Further analysis of VAE-derived feature trajectories showed significant temporal patterns. ECG markers that increased markedly in the dementia-develop group as the index date approached were associated with QT interval differences and T-wave morphology variations. Conversely, markers that increased in the dementia-free group but decreased in the dementia-develop group were linked to ST wave differences and amplitude disparities (Figure1). These findings underscore the potential of VAE-derived ECG features to distinguish between dementia-develop and dementia-free individuals over time. Conclusion: The model using VAE-derived ECG features highlighted the potential of combining ECG data with advanced machine learning techniques for early dementia screening.