Latent representation of intracardiac electrograms using an AI-based autoencoder for rhythm classification

Abstract Background Intracardiac electrograms (EGMs) contain crucial information about cardiac electrical activity, yet their high dimensionality can hinder interpretation. AI-based autoencoders learn low dimensional representation of these signals, facilitating arrhythmia and sinus rhythm classification, for example. We investigated how a compact autoencoder-derived latent space might enhance data interpretability and improve rhythm discrimination in intracavitary unipolar and bipolar EGMs. Purpose We aimed to develop and evaluate an AI-based autoencoder that compresses unipolar and bipolar EGM signals into a concise latent space. Specifically, we sought to differentiate atrial fibrillation (AF) from paced sinus rhythms at 300 ms (SR300) and 600 ms (SR600), while preserving key diagnostic features. We also highlight the potential of this approach for broader applications, such as patient-specific analyses or anatomical segmentation. Methods We collected unipolar and bipolar EGMs from 28 AF patients at 500 Hz for 2.5 seconds using a multi-electrode catheter with an electroanatomical mapping system, resulting in 1,250 samples per signal. In total, we employed 479,042 signals for each EGM type (186,704 AF, 163,353 SR300, and 128,985 SR600), averaging 17,108±5,984 signals per patient. We then trained an AI-based autoencoder in an unsupervised manner to capture a 64-dimensional latent representation of these signals (Figure 1). For unipolar signals, the test loss was 0.0194 and the test R² was 0.8763, and for bipolar signals, the test loss was 0.0292 and the test R² was 0.6364. After training, we applied t-SNE to visualise how effectively the latent space separates different rhythms (Figure 2). Next, we used the latent features as inputs to classification models, specifically a single-layer MLP, to evaluate their ability to distinguish among the three cardiac rhythms (Figure 2). Results We observed clear rhythm separation in the 64-dimensional latent space when visualised with t-SNE. The MLP classifier achieved high F1-scores for unipolar EGMs (0.9856 for AF, 0.9732 for SR300, 0.9765 for SR600) with an overall accuracy of 0.9789. For bipolar EGMs, the MLP yielded F1-scores of 0.9900 for AF, 0.9858 for SR300, and 0.9807 for SR600, with an overall accuracy of 0.9861. Conclusions Our findings demonstrate that an AI-based autoencoder effectively compresses intracardiac EGMs into a clinically meaningful latent representation, enabling highly accurate arrhythmia classification while improving interpretability. By focusing on the most relevant features, this approach can support broader clinical applications, such as individualised arrhythmia management, real-time risk stratification, and improved guidance for interventions (e.g., ablation). Integrating patient-specific factors and anatomical data into this framework may further optimise diagnostic accuracy and personalise treatment strategies.Autoencoder Results