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

TL;DR

ExChanGeAI delivers an end-to-end, open-source platform for ECG analysis that streamlines data handling, visualization, and privacy-preserving fine-tuning while enabling cross-instance model exchange via ONNX. The authors introduce CardX, a lightweight ECG foundation model based on a Mixture-of-Architectures design trained on diverse, large-scale ECG datasets, and demonstrate its robust performance and efficiency across multiple external test sets. Across intra- and inter-dataset evaluations, CardX often achieves top or near-top F1 scores with substantially fewer parameters and FLOPs than heavier foundation models, highlighting the importance of empirical validation in data-limited clinical settings. The work emphasizes open science and reproducibility, showing that end-to-end platforms can democratize access to advanced ECG analysis without sacrificing performance or privacy, though it also notes limitations related to data quality, infrastructure, and domain shift.

Abstract

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 .

Figures (5)

• Figure 1: Analyse (left) and Finetuning (right) page of ExChanGeAI with sample data.
• Figure 2: Overview of the end-to-end platform ExChanGeAI and its three main distinct parts: 1. Analysis, 2. The AI Ecosystem and 3. Interoperability. All parts are transparent to the user, and interact without manual interaction.
• Figure 3: Overview of the semi-automatic fine-tuning process. The user provides labels and 12-lead ECGs, and can change the default configuration if necessary. The backend process is executed without requiring any previous programming knowledge.
• Figure 4: Mixture of Architectures scheme of the electrocardiogram foundation model. The encoders are sparsely activated (dotted line) for each Electrocardiogram, depending on the router. CardX supports self- and semi-supervised training approaches. The architecture is structured for multi-task learning, capable of addressing multiple objectives: multivariate time-aligned ECG reconstruction, corrected QT-time (QTc) regression, origin dataset prediction, and prevalent feature prediction. For the final foundation model intended for distribution, the reconstruction branch is intentionally removed, and replaced by a linear layer, the classification head. This architectural decision is justified by the pre-training role of reconstruction, its absence of direct clinical utility in downstream tasks, and the resulting advantages in model size reduction and privacy preservation.
• Figure S1: Flowchart of the pre-processing applied to any ECG data while being loaded into the application, independent of the file format.