Interpretability and accuracy of machine learning algorithms for biomedical time series analysis – a scoping review

Current research in biomedical time series (BTS) (e.g., ECG, EEG) analysis focuses on applications of various deep learning approaches to improve classification, prediction, or assessment of states and disorders. When trained on sufficiently large datasets, such approaches mostly lead to highly accurate, yet uninterpretable models, sometimes with a possibility for post-hoc explainability. Since high-stake areas such as healthcare warrant model explanations and, where possible, high interpretability in addition to model efficiency, there is nowadays a surprising scarcity of interpretable machine learning models proposed in this field. Although the machine learning community is aware of the need for interpretable machine learning in BTS analysis, the proposed models do not reflect this need. In this scoping review, we considered over 30,000 studies from the Web of Science database, screened nearly 500 studies, and selected over 50 high-quality studies for detailed analysis. These studies focus on interpretable methods, accurate methods, and approaches bridging the two. Most studies analyzed ECG and EEG signals and concentrated on a limited range of applications, including emotion recognition, heart diseases, epilepsy, and motor imagery, reflecting the scarcity of quality public datasets. K-nearest neighbors and decision trees were the most used interpretable methods, while convolutional neural networks with recurrent or attention layers, achieved the highest accuracy. The methods that balance interpretability and accuracy in BTS analysis include advanced generalized additive models and optimization-based approaches for decision trees, rule learning, and linear models. These approaches warrant further studies, as only a few of them were applied in BTS analysis.