Deep Embedded Clustering for Data-Driven ECG Exploration Using Continuous Wavelet Transforms

Due to the length, complexity, and inter-subject variation of physiological signals acquired non-invasively, datadriven analysis is increasingly valuable in Smart Health, precision medicine, and in studying physiological dynamics. Knowledge discovery in medical research requires that domain experts analyze complex, high-dimensional data and signals, which is facilitated by dimensionality reduction and clustering. The current paper presents an unsupervised machine learning approach to clustering time-frequency features from ECG records using deep embedded clustering, which optimizes a clustering metric that maps high-dimensional time-frequency data to a lower dimensional latent space. These clusters can be obtained in arbitrarily low dimensions, and are subsequently analyzed with visual analytics to uncover structures and patterns. This technique is applied to time segments of continuous wavelet transforms of ECG records, representing a variety of conditions. Preliminary results on publicly-available ECG records indicate that deep embedded clustering produces a low-dimensional learned representation of time-frequency characteristics that facilitates signal exploration and improves interpretability.