ExChanGeAI: An End-to-End Platform and Efficient Foundation Model for Electrocardiogram Analysis and Fine-tuning

Electrocardiogram data, one of the most widely available biosignal data, has become increasingly valuable with the emergence of deep learning methods, providing novel insights into cardiovascular diseases and broader health conditions. However, heterogeneity of electrocardiogram formats, limited access to deep learning model weights and intricate algorithmic steps for effective fine-tuning for own disease target labels result in complex workflows. In this work, we introduce ExChanGeAI, a web-based end-to-end platform that streamlines the reading of different formats, pre-processing, visualization and custom machine learning with local and privacy-preserving fine-tuning. ExChanGeAI is adaptable for use on both personal computers and scalable to high performance server environments. The platform offers state-of-the-art deep learning models for training from scratch, alongside our novel open-source electrocardiogram foundation model CardX, pre-trained on over one million electrocardiograms. Evaluation across three external validation sets, including an entirely new testset extracted from routine care, demonstrate the fine-tuning capabilities of ExChanGeAI. CardX outperformed the benchmark foundation model while requiring significantly fewer parameters and lower computational resources. The platform enables users to empirically determine the most suitable model for their specific tasks based on systematic validations.The code is available at https://imigitlab.uni-muenster.de/published/exchangeai .

💡 Research Summary

The paper introduces ExChanGeAI, a web‑based end‑to‑end platform that integrates every step required for modern electrocardiogram (ECG) analysis, from loading heterogeneous file formats to visualizing waveforms, preprocessing signals, training deep‑learning models, fine‑tuning them on user‑defined disease labels, and finally deploying the resulting models. The platform supports a wide range of common ECG formats (DICOM, EDF, CSV, etc.), automatically applies a configurable preprocessing pipeline (moving‑median denoising, min‑max scaling, Z‑score normalization), and offers four interactive visualisation modes: raw time‑series, QRS complexes, fiducial‑point annotations, and time‑aligned median beats. Users can inspect and correct annotations directly in the browser, lowering the barrier for clinicians without programming expertise.

Model management is built around the Open Neural Network Exchange (ONNX) standard, allowing the same model file to run on a personal computer, a high‑performance server, or an edge device. Trained models are stored and shared via a WebDAV‑based “Model ExChanGe” server that supports versioning and access control, facilitating collaborative research. The platform also includes privacy‑preserving learning options such as federated and swarm learning, enabling multi‑institutional training without moving raw ECG data.

The core technical contribution is CardX, a novel ECG foundation model pre‑trained on more than one million 12‑lead recordings collected from six public databases and one proprietary source. CardX adopts a lightweight hybrid architecture that combines convolutional blocks with a transformer‑style temporal attention mechanism. It contains roughly 15 million parameters and requires about 201 million floating‑point operations per inference (MFLOPs), which is roughly 70‑fold fewer FLOPs and six‑times fewer parameters than the previously published ECG‑FM model (≈90 M parameters, 14 GFLOPs). Pre‑training employed a mix of supervised learning on labelled data and semi‑supervised contrastive‑masking on unlabelled recordings, thereby leveraging both label‑rich and label‑poor data.

To evaluate the platform and CardX, the authors fine‑tuned several architectures (XceptionTime, InceptionTime, DSAIL SNU, ECG‑FM, and CardX) on nine diagnostic tasks covering myocardial infarction, ST/T changes, conduction disturbances, hypertrophy, bundle‑branch block sub‑types, and a clinically relevant “re‑vascularization needed” decision that is not present in the PTB‑XL benchmark. Training and testing were performed on the internal PTB‑XL dataset and three external datasets: the Yang et al. collection, MIMIC‑IV‑ECG, and a newly curated Emergency Department Münster (EDMS) set. Performance metrics included weighted F1‑score, inter‑quartile range (IQR), coefficient of variation (CV), as well as model size and computational cost.

CardX achieved an average weighted F1 of 0.611 and a median weighted F1 of 0.645 across the external test sets, with the lowest IQR (0.123) and CV (0.358) among all pre‑trained models, indicating superior robustness to dataset shift. While XceptionTime and InceptionTime obtained slightly higher peak F1 scores on some tasks, they required considerably more parameters (≈0.4 M) and FLOPs (≈460 MFLOPs for InceptionTime, 256 MFLOPs for XceptionTime). In data‑constrained scenarios, CardX outperformed the larger ECG‑FM (90 M parameters, 14 GFLOPs) and the smaller DSAIL SNU (2 M parameters, 89 MFLOPs) both in accuracy and stability, demonstrating that a well‑designed lightweight foundation model can rival much larger networks when fine‑tuned on limited target data.

The platform’s “one‑click fine‑tuning” feature allows users to upload a label file, select a pre‑trained backbone, choose whether to freeze all layers or only the classification head, and launch a full training‑validation‑test cycle without writing code. Hyper‑parameters such as learning rate, batch size, and number of frozen layers are exposed via simple sliders. Results are presented instantly as bar charts and confusion matrices, enabling rapid iteration.

Limitations noted by the authors include the current focus on 12‑lead ECGs, which restricts applicability to single‑lead wearable devices, and a reliance on high‑quality manual labels; noisy or mis‑labelled data can degrade fine‑tuning performance. Moreover, the pre‑training corpus is dominated by recordings from Western hospitals, raising concerns about demographic bias when applying CardX to under‑represented populations. Future work will explore extending the platform to low‑channel signals, incorporating automated label‑cleaning pipelines, and augmenting the pre‑training set with diverse global datasets to improve fairness.

In summary, ExChanGeAI delivers a user‑friendly, scalable, and privacy‑aware environment for ECG deep‑learning research, while CardX provides an efficient, open‑source foundation model that balances performance, computational demand, and robustness across heterogeneous clinical datasets. The combination of these tools promises to lower the barrier for clinicians and researchers to develop, evaluate, and share AI‑driven ECG diagnostics, accelerating translation from academic prototypes to real‑world clinical decision support.