Abstract 4146665: Unveiling the Biological Subtypes of Left Ventricular Diastolic Dysfunction through Unsupervised Clustering of ECG Data

Left Ventricular Diastolic Dysfunction (LVDD) presents a challenge in terms of management given its multifaceted etiology and complex presentation. Conventional diagnostic categories may fail to fully reflect underlying pathophysiology variability. Using unsupervised clustering on electrocardiogram (ECG) data is a novel approach to capture subtle changes in the electrical signatures seen in ECGs and help uncover naturally occurring patient subtypes that may be invisible to traditional echocardiographic (echo) assessment. We performed Natural Language Processing on long-form echo reports to extract annotations of LVDD. Echo reports were paired with an ECG performed within 7 days of the echo. Patients were excluded if their ECG diagnostic codes indicated conduction abnormalities, concurrent ischemia, and prior intervention. We applied a Convolutional Autoencoder to compress high-dimensional ECG data into a 128-dimensional, feature-rich space suitable for clustering. K-means clustering was then utilized to categorize patients into groups based on transformed ECG features. Silhouette analysis was employed to determine the optimal number of clusters and quantify differences between groups. Chi-square and ANOVA tests examined for differences in age, sex, and race across clusters. We identified 22,183 patients paired with 96,921 ECGs with clinician confirmed LVDD. Silhouette score analysis on K-Means clustering identified 2 as the optimal number of clusters (Figure 1). Labels associated with each data point were visualized in 2D and 3D projections with the t-Stochastic Neighbor Embedding technique (Figure 2). Identified clusters differed significantly in terms of age, sex, and race, with p-values of <0.00001 for each demographic. The identification of two distinct clusters underscores the potential of ECG-based phenotyping in revealing new biological subtypes of LVDD. Extending this technique to integrate other clinical variables and longitudinal data may enhance the clinical utility of identified subtypes. Future work will look at cluster-wise distribution of co-morbidities and longitudinal outcomes. Applying similar methodologies to other cardiac disorders may improve our approach to their diagnosis and management. Employing unsupervised clustering on ECG has proven to be effective in identifying novel subtypes of LVDD, showcasing the potential of ECG data to refine diagnostic and therapeutic strategies.