ecgXfusion - Innovative ECG analysis based on trustworthy AI

The novel AI-based model architecture xECGArch® comprises two separate Convolutional Neural Networks (CNN), which are distinguished by a differential dimensioning of the receptive fields and, consequently, a differential observation horizon. This enables the models to learn different features in the electrocardiogram (ECG) and utilize them for classification. The short-term CNN learns beat morphological features, while the long-term CNN learns rhythm. The features correspond to the characteristics typically used in ECG analysis, which enables medical professionals to interpret the features.

© M. Goettling et al. (2024) [CC BY 4.0]

Publications:

• M. Goettling, A. Hammer, H. Malberg, and M. Schmidt, “xECGArch: a trustworthy deep learning architecture for interpretable ECG analysis considering short-term and long-term features,” Sci Rep, vol. 14, no. 1, p. 13122, Jun. 2024, doi: 10.1038/s41598-024-63656-x.

• A. Hammer, M. Goettling, H. Malberg, A. Linke, and M. Schmidt, “An explainable AI for trustworthy detection of atrial fibrillation on reduced lead ECGs in mobile applications,” European Heart Journal, vol. 45, no. Supplement_1, Oct 2024, doi: 10.1093/eurheartj/ehae666.3497.

• A. Hammer, H. Malberg, and M. Schmidt, “Morphology Features Self-Learned by Explainable Deep Learning for Atrial Fibrillation Detection Correspond to Fibrillatory Waves,” in 2024 Computing in Cardiology (CinC), pp. 1–4., doi: 10.22489/CinC.2024.305.

• A. Hammer, F. Hänsch, L. Seliger, H. Malberg, A. Linke, and M. Schmidt, „Explainable and interpretable AI visualises self-learned clinically relevant ECG characteristics of rhythm and morphology paving the way for trustworthy diagnostic support“, European Heart Journal, vol. 46, no. Supplement_1, Nov. 2025, doi: 10.1093/eurheartj/ehaf784.4412.
• A. Hammer, M. Zannini, H. Malberg, and M. Schmidt, „Transfer Learning to Focus Self -Learning AI on Rhythm Improves Interpretability in Atrial Fibrillation Detection“, in Computing in Cardiology 2025, São Paulo, Brazil, Sep. 2025, S. 1–4. doi: 10.22489/CinC.2025.396.

xFuseMap® represents the explanations of two models in a merged saliency map. Saliency maps are a form of representation in which the relevance of data points for the classification decision of a model is displayed by coloring the signal. xFuseMap® enables the interpretable and comprehensible visualization of features of different models and feature domains. Each model is assigned a unique color. Data points that are relevant for both or none of the models are colored in mixed colors, creating a two-dimensional color space with the models on one axis each.

xFuseMap® enables the integrated visualization of the explanations derived from both xECGArch® models (long and short term) and their corresponding feature domains (rhythm and morphology) within an electrocardiogram (ECG). Morphology and rhythm are fundamental ECG characteristics that are utilized by physicians for diagnostic purposes and are therefore amenable to classification. The objective is to equip medical professionals with the most reliable diagnostic support.

© A. Hammer

Publication:

• A. Hammer, M. Göttling, H. Malberg, A. Linke, S. Richter, N. Mangner and M. Schmidt, “Fusion of automatically learned rhythm and morphology features matches diagnostic criteria and enhances AI explainability,” npj Artificial Intelligence 1, 19 (2025). https://doi.org/10.1038/s44387-025-00022-w..

• A. Hammer, F. Hänsch, L. Seliger, H. Malberg, A. Linke, and M. Schmidt, „Explainable and interpretable AI visualises self-learned clinically relevant ECG characteristics of rhythm and morphology paving the way for trustworthy diagnostic support“, European Heart Journal, vol. 46, no. Supplement_1, Nov. 2025, doi: 10.1093/eurheartj/ehaf784.4412.

The autonomic nervous system (ANS) controls numerous functions of the body and shows specific reactions depending on the individual's condition. The evaluation of the ANS therefore allows far-reaching statements to be made, for example on the risk of suffering life-threatening illnesses, an evaluation of sleep and the recording of stress levels.

The Two-Dimensional Signal Warping (2DSW) technique offers the possibility to record complex morphological changes between heartbeats and to make such statements non-invasively (e.g. by analyzing QT variability). The aim is to significantly improve diagnostic possibilities, in particular prediction and prevention, without causing additional stress for patients.

Further information can be found at 2dsw.com.

© Martin Schmidt [CC BY-SA 4.0]

Publications:

• M. Schmidt, M. Baumert, A. Porta, H. Malberg, and S. Zaunseder, “Two-Dimensional Warping for One-Dimensional Signals—Conceptual Framework and Application to ECG Processing,” IEEE Transactions on Signal Processing, vol. 62, no. 21, pp. 5577–5588, Nov. 2014, doi: 10.1109/TSP.2014.2354313.

• M. Schmidt, M. Baumert, H. Malberg, and S. Zaunseder, “Iterative two-dimensional signal warping—Towards a generalized approach for adaption of one-dimensional signals,” Biomedical Signal Processing and Control, vol. 43, pp. 311–319, May 2018, doi: 10.1016/j.bspc.2018.03.016.

• For further publications, visit 2dsw.com