The causes of ventricular fibrillation (VF) are not yet elucidated, and it has been proposed that different mechanisms might exist. Moreover, conventional analysis methods do not seem to provide time or frequency domain features that allow for recognition of different VF patterns in electrode-recorded biopotentials. The present work aims to determine whether low-dimensional latent spaces could exhibit discriminative features for different mechanisms or conditions during VF episodes. For this purpose, manifold learning using autoencoder neural networks was analyzed based on surface ECG recordings. The recordings covered the onset of the VF episode as well as the next 6 min, and comprised an experimental database based on an animal model with five situations, including control, drug intervention (amiodarone, diltiazem, and flecainide), and autonomic nervous system blockade. The results show that latent spaces from unsupervised and supervised learning schemes yielded moderate though quite noticeable separability among the different types of VF according to their type or intervention. In particular, unsupervised schemes reached a multi-class classification accuracy of 66%, while supervised schemes improved the separability of the generated latent spaces, providing a classification accuracy of up to 74%. Thus, we conclude that manifold learning schemes can provide a valuable tool for studying different types of VF while working in low-dimensional latent spaces, as the machine-learning generated features exhibit separability among different VF types. This study confirms that latent variables are better VF descriptors than conventional time or domain features, making this technique useful in current VF research on elucidation of the underlying VF mechanisms.

In recent years, a wide variety of Machine Learning (ML) algorithms, including Deep Learning (DL) methods, have been proposed for electrocardiogram (ECG) beat classification. However, accurately discerning ECG beat types faces challenges due to noise interference and inherent imbalances among different classes. Moreover, understanding mathematical models enclosed by black-box learning systems is an open issue today. Our study employed a manifold learning algorithm capable of mapping high-dimensional data into a latent space to conduct a comprehensive analysis within a neural network learning framework. This approach involved the following studies: (1) Examining the intermediate high-dimensional latent space in simple architectures by studying its projection into a visualizable latent space; (2) Exploring the influence of class imbalance on the configuration of the latent space; (3) Evaluating and analyzing the compensatory effects of employing diverse DL architectures, such as modified autoencoders and Generative Adversarial Networks (GANs), specifically in generating data augmentation through synthetic beats. The experimental results demonstrated the effectiveness of our methodology in mitigating noise and addressing inter-class imbalance, notably enhancing the diagnostic Area Under the ROC Curve (AUC) in ECG signal analysis. Implementation of GAN data augmentation techniques resulted in a 2% improvement, elevating the AUC from 0, 9332 to 0, 9520 in the biclass dataset. Similarly, the AUC values increased by 2%, from 0, 9020 to 0, 9223, for the multiclass dataset. These findings highlight the impact of appropriate data augmentation techniques on AUC improvement. Furthermore, visualizing latent spaces during beat classifier design significantly contributes to the development of solid and principled multiclass beat-discriminating systems.

T-wave alternans (TWA) is a biomarker for sudden cardiac death prediction, characterized by subtle variations in the amplitude or morphology of consecutive T-waves in electrocardiographic studies. Electrocardiographic imaging (ECGI) offers increased spatial resolution, enabling TWA distribution analysis across the epicardium. However, existing TWA estimation methods disregard ECGI spatial information by analyzing each signal independently. To address this gap, we present a novel, subject-specific, interpretable manifold learning-based TWA estimation method tailored to ECGI. First, Uniform Manifold Approximation and Projection (UMAP) reduces input data dimensions. Second, the Louvain algorithm detects communities and identifies the TWA-dominant community. Finally, the location of this community and the rest of the communities is compared, and a Bootstrap-based TWA classifier is applied. A customized Shapley additive explanations method was developed to identify the signal segments most affecting the algorithm decisions to enhance explainability. Reducing the input data to 18 dimensions improved the separation of the TWA-dominant community, with an average normalized distance of 0.28. The Bootstrap analysis showed that the TWA-dominant community had a distance metric up to 0.2 above the confidence interval upper limit. The TWA-dominant community input signals showed different TWA patterns, namely, hump-shaped and amplitude-shifted TWA, and the interpretability algorithm revealed that UMAP focuses on them when projecting points into the latent space. Our method achieved maximum accuracy in subjects with known outcomes and made consistent patient decisions based on input signals. This study introduces the first ECGI-specific TWA detection method. Its subject-specific nature enables the extraction of individual-specific characteristics, offering personalized diagnostic insights.