Manifold learning for personalized and label-free detection of cardiac arrhythmias

Abstract

Electrocardiograms (ECGs) provide non-invasive measurements of heart activity and are well-established tools for detecting cardiac arrhythmias. Although supervised machine learning has emerged as a promising approach for automated heartbeat classification, substantial variations in ECG signals across individuals and leads, combined with inconsistent labeling standards and dataset biases, make it difficult to develop generalizable models. Dimensionality reduction is an unsupervised approach that maps high-dimensional data into a lower-dimensional space while preserving the underlying structure, enabling data visualization and pattern discovery. Conventional methods, e.g., principal component analysis, prioritize large (and often obvious) variances in the data and typically overlook subtle yet clinically relevant patterns. Here, we show that nonlinear dimensionality reduction (NLDR) algorithms, e.g., t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP), can identify medically relevant features in ECG signals without pretraining or prior information. Using the MIT-BIH Arrhythmia Database, we show that: a) applying NLDR to a mixed population of heartbeats reveals inter-individual differences in morphology, as signals from the same person cluster together in latent spaces; and b) applying NLDR to the heartbeats of a single individual separates normal beats from arrhythmias into distinct clusters, which can be identified by clustering algorithms in an unsupervised manner. To our knowledge, this is the first systematic evaluation of NLDR for unsupervised arrhythmia detection. As a proof-of-concept, we demonstrate that both UMAP and t- SNE achieved trustworthiness scores of ≥0.95, indicating that local neighborhoods from the original space are well preserved in the embedding. Moreover, classification on the 2D embeddings outperforms classification in the original high-dimensional space, with a simple k-NN classifier discriminating individual recordings with ≥80% accuracy and identifying arrhythmias with a median accuracy of ≥98% and median F1-score of ≥85%. These results show that NLDR holds much promise for cardiac monitoring, and for personalized healthcare beyond cardiology.