Benchmarking Approaches: Time Series Versus Feature-Based Machine Learning in ECG Analysis on the PTB-XL Dataset.

Extensive research has been conducted on time series and tabular data in the context of classification tasks, considering their distinct data domains. While feature extraction enables the transformation of series into tabular data, direct comparative comparisons between these data types remain scarce. Especially in the domain of medical data, such as electrocardiograms (ECGs), deep learning faces challenges due to its lack of easy and fast interpretability and explainability. However, these are crucial aspects for a wide and reliable adoption in the field. In our study, we assess the performance of XGBoost and InceptionTime on ECG features and time series data respectively. Our findings reveal that features extracted from ECG signals not only achieve competitive performance but also retain advantages during training and inference. These advantages encompass accuracy, resource efficiency, stability, and a high level of explainability.

Assessing the Reliability of Machine Learning Explanations in ECG Analysis Through Feature Attribution.

Feature attribution methods stand as a popular approach for explaining the decisions made by convolutional neural networks. Given their nature as local explainability tools, these methods fall short in providing a systematic evaluation of their global meaningfulness. This limitation often gives rise to confirmation bias, where explanations are crafted after the fact. Consequently, we conducted a systematic investigation of feature attribution methods within the realm of electrocardiogram time series, focusing on R-peak, T-wave, and P-wave. Using a simulated dataset with modifications limited to the R-peak and T-wave, we evaluated the performance of various feature attribution techniques across two CNN architectures and explainability frameworks. Extending our analysis to real-world data revealed that, while feature attribution maps effectively highlight significant regions, their clarity is lacking, even under the simulated ideal conditions, resulting in blurry representations.

Post Hoc Sample Size Estimation for Deep Learning Architectures for ECG-Classification.

Deep Learning architectures for time series require a large number of training samples, however traditional sample size estimation for sufficient model performance is not applicable for machine learning, especially in the field of electrocardiograms (ECGs). This paper outlines a sample size estimation strategy for binary classification problems on ECGs using different deep learning architectures and the large publicly available PTB-XL dataset, which includes 21801 ECG samples. This work evaluates binary classification tasks for Myocardial Infarction (MI), Conduction Disturbance (CD), ST/T Change (STTC), and Sex. All estimations are benchmarked across different architectures, including XResNet, Inception-, XceptionTime and a fully convolutional network (FCN). The results indicate trends for required sample sizes for given tasks and architectures, which can be used as orientation for future ECG studies or feasibility aspects.

The Necessity of Multiple Data Sources for ECG-Based Machine Learning Models.

Even though the interest in machine learning studies is growing significantly, especially in medicine, the imbalance between study results and clinical relevance is more pronounced than ever. The reasons for this include data quality and interoperability issues. Hence, we aimed at examining site- and study-specific differences in publicly available standard electrocardiogram (ECG) datasets, which in theory should be interoperable by consistent 12-lead definition, sampling rate, and measurement duration. The focus lies upon the question of whether even slight study peculiarities can affect the stability of trained machine learning models. To this end, the performances of modern network architectures as well as unsupervised pattern detection algorithms are investigated across different datasets. Overall, this is intended to examine the generalization of machine learning results of single-site ECG studies.