🤖 machine learning

Domain-Adapted Fine-Tuning of ECG Foundation Models for Multi-Label Structural Heart Disease Screening

This paper demonstrates that the most effective strategy for screening multiple structural heart diseases via ECG is to combine in-domain self-supervised adaptation with selective supervised fine-tuning of a pretrained ECG foundation model.

Original authors: Duc N. Do, Minh N. Do, Dang Nguyen, Khanh T. Q. Le, Khoa D. Pham, Hung N. Huynh, Phi Pham-Van-Hoang, Quan K. Huynh, Ramez M. Odat, Perisa Ashar, Ethan Philip Lowder, Minh H. N. Le, Hoang Le, Phat V. H

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Duc N. Do, Minh N. Do, Dang Nguyen, Khanh T. Q. Le, Khoa D. Pham, Hung N. Huynh, Phi Pham-Van-Hoang, Quan K. Huynh, Ramez M. Odat, Perisa Ashar, Ethan Philip Lowder, Minh H. N. Le, Hoang Le, Phat V. H. Nguyen, Quan Le, Jacques Kpodonu, Phat K. Huynh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Teaching a "Musical Ear" to Spot Heart Problems

Imagine you have a world-class musician who has spent years listening to every single song ever recorded. They can tell you the rhythm, the key, and the genre of almost anything just by hearing a few seconds of it. This musician is like an ECG Foundation Model—a powerful AI that has "listened" to millions of heartbeats (ECG signals) and learned the "music" of the human heart.

However, there is a catch: this musician is an expert in rhythm (is the heart beating too fast or too slow?), but they aren't necessarily an expert in structural engineering (is the heart's "room" too small, or is one of its "valves" leaking?).

The Problem:
To find out if someone has serious structural heart disease (SHD), doctors usually need an echocardiogram—an ultrasound of the heart. It’s the "gold standard," but it’s expensive, requires special equipment, and takes a lot of time. Doctors want a way to use the cheap, quick, and everywhere-available ECG (the little stickers on your chest) to act as a "triage" system—a way to flag people who really need that expensive ultrasound.

The Research Question:
Can we take that "world-class musician" AI and quickly teach them to listen for the subtle "cracks" and "thumps" in the music that signal a structural problem?

The Experiment: The "Training Camp"

The researchers didn't just ask the AI to guess. They put it through a specific training regimen to see which method worked best. Think of it like training an athlete:

The "Old School" Way (Baseline A): This is like giving a coach a checklist of specific things to look for (like "is the beat steady?"). It works, but it's limited.
The "Start from Scratch" Way (Baseline B): This is like taking a baby and trying to teach them music from zero. It takes a massive amount of effort and data.
The "Quick Study" Way (The Researchers' Method): This is the star of the show. They took the "expert musician" and gave them two specific lessons:
- Lesson 1 (Self-Supervised Adaptation): They let the AI listen to thousands of new heartbeats without telling it what they meant. This helped the AI get used to the "accent" and "dialect" of the specific hospital's data.
- Lesson 2 (Selective Fine-Tuning): Instead of trying to rewire the AI's entire brain (which is expensive and slow), they only updated the "top layers"—the parts of the brain responsible for making final decisions.

The Results: Finding the "Sweet Spot"

The researchers discovered that you don't need to rebuild the whole brain to learn a new skill.

The "Goldilocks" Effect: If you change too little of the AI, it doesn't learn the new task. If you change too much (full fine-tuning), it becomes "overwhelmed" and loses its original expertise. But if you change just the right amount (the "partial fine-tuning" they tested), it becomes incredibly accurate at spotting things like leaky valves or thickened heart walls.
The "Less is More" Winner: They found a specific setting (called $b=5$ ) that was like a "smart athlete." It wasn't the biggest or the heaviest model, but it was the most efficient. It was excellent at catching the most important cases without needing a supercomputer to run it.
The "Extra Info" Myth: Interestingly, they tried giving the AI extra data, like the patient's age or sex. Surprisingly, the AI was so good at "hearing" the heart's problems in the waveform itself that the extra written info didn't actually help it much. The "music" told the whole story.

Why This Matters (The "So What?")

In the real world, this could change how we handle heart health.

Imagine walking into a pharmacy or a quick-check clinic. You get a simple, 10-second ECG. An AI—trained using this "selective" method—listens to your heartbeat. It doesn't give you a final diagnosis, but it says, "Hey, I hear a subtle 'thump' that sounds like a valve might be leaking. You should go see a specialist for an ultrasound."

In short: This paper proves we can take powerful, general AI and "sharpen" it into a highly specialized medical tool very efficiently, helping doctors find the right patients at the right time.

Technical Summary: Domain-Adapted Fine-Tuning of ECG Foundation Models for Multi-Label Structural Heart Disease Screening

1. Problem Statement

Structural Heart Disease (SHD)—including ventricular dysfunction, hypertrophy, and valvular diseases—is typically diagnosed via transthoracic echocardiography (TTE). However, TTE is expensive, resource-intensive, and requires specialist availability, making it unsuitable as a universal first-line screening tool. While 12-lead electrocardiograms (ECGs) are ubiquitous and low-cost, most existing AI models trained on ECGs focus on rhythm or conduction abnormalities rather than the structural phenotypes confirmed by echocardiography.

The core research question is: How can open-source pretrained ECG Foundation Models (FMs) be most effectively adapted to detect multi-label, echo-confirmed SHD while balancing predictive performance and computational adaptation costs?

2. Methodology

The study utilizes the EchoNext Mini-Model dataset (a PhysioNet release), containing 100,000 de-identified 12-lead ECGs paired with echocardiographic labels. The researchers focused on six moderate-or-greater SHD endpoints:

Reduced LVEF ( $\le 45\%$ )
Increased LV wall thickness ( $\ge 1.3$ cm)
Aortic stenosis
Mitral regurgitation
Tricuspid regurgitation
Right ventricular (RV) systolic dysfunction

Experimental Pipeline:

Baselines: The study compared engineered ECG features (using XGBoost), end-to-end waveform learning from scratch (ResNet1D), and supervised probing of a frozen, in-domain adapted ECG-FM.
Primary Transfer Strategy: A two-stage adaptation process:
1. Stage 1 (In-domain SSL): Continued self-supervised learning (using wav2vec 2.0-style masking) on the EchoNext waveforms to align the pretraining distribution with the target clinical population.
2. Stage 2 (Selective Supervised Fine-Tuning): Updating a specific number of upper transformer blocks ( $b \in \{1, 5, 9, 12\}$ ) while keeping the convolutional extractor frozen.
Ablations & Extensions: The authors investigated Low-Rank Adaptation (LoRA), late fusion of tabular covariates (age, sex, etc.), Mixture-of-Foundation-Models (MoFM), and task-decomposed binary training (one model per disease).

3. Key Contributions

Systematic Evaluation of Transfer Learning: Unlike previous studies that focus on task-specific supervised learning, this paper rigorously disentangles the effects of domain alignment (SSL) and adaptation depth (fine-tuning) in ECG FMs.
Identification of Optimal Adaptation Depth: The study identifies that "full" fine-tuning is not always superior; rather, selective updates to the transformer layers provide a better performance-to-parameter ratio.
Benchmark for SHD Detection: It provides a standardized protocol for evaluating how foundation models can be repurposed for complex, multi-label clinical triage tasks.

4. Results

Superiority of Adapted FMs: The adapted ECG-FM models outperformed all other configurations. The best performing model (partial fine-tuning with $b=9$ blocks) achieved a macro-AUROC of 0.8509 and macro-AUPRC of 0.4261.
The Efficiency "Sweet Spot": A more parameter-efficient configuration ( $b=5$ blocks) maintained a nearly identical AUROC (0.8501) but achieved the highest fixed-threshold macro-F1 (0.3691), suggesting it is highly effective for practical clinical decision-making.
Diminishing Returns of Complexity:
- Full Model Fine-Tuning: Updating the convolutional extractor did not improve performance over updating only the transformer blocks.
- Late Fusion: Adding tabular covariates (like age or sex) did not improve threshold-independent discrimination (AUROC/AUPRC), implying the waveform encoder already captured this information.
- MoFM & LoRA: Mixture-of-Models and LoRA did not surpass the performance of the best single-backbone selective fine-tuning approach.
Multi-label vs. Binary: The unified multi-label model outperformed an ensemble of six independent binary classifiers in both AUROC and F1, suggesting that shared representation learning benefits the detection of correlated structural abnormalities.

5. Significance

This research demonstrates that ECG Foundation Models are highly viable for echocardiography triage. By using a two-stage approach—aligning the model to the specific data distribution via self-supervised learning and then selectively fine-tuning the transformer layers—clinicians can implement highly accurate, low-cost screening tools. This has the potential to optimize healthcare workflows by prioritizing high-risk patients for confirmatory echocardiography, thereby reducing costs and improving the early detection of structural heart disease.