📄 cardiovascular medicine

Leveraging the wearable 1-lead ECG signal: From cardiac rhythm to cardiac function assessment

This study demonstrates that explainable deep learning models can accurately predict left ventricular function from wearable 1-lead ECGs with performance nearly matching 12-lead ECGs, thereby enabling broader clinical applications and enhanced patient autonomy in remote or resource-limited settings.

Original authors: van der Valk, V. O., Atsma, D., Scherptong, R., Staring, M.

Published 2026-02-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: van der Valk, V. O., Atsma, D., Scherptong, R., Staring, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: From "Heartbeat" to "Heart Health"

Imagine your heart is a house. For a long time, doctors have used a 12-lead ECG as a high-tech security system with 12 different cameras placed all around the house. This gives them a complete, 360-degree view of everything happening inside, allowing them to spot even the smallest structural problems. However, you can only use this system if you go to the hospital, and it requires a technician to set it up.

On the other hand, we now have smartwatches that act like a single, tiny security camera you wear on your wrist. It's great for watching the "rhythm" of the house (is the door opening and closing at the right speed?), but because it only has one lens, it was thought to be useless for checking the structural integrity of the house (like a weak wall or a damaged foundation).

This study asked a simple question: Can that single, tiny camera on the wrist actually tell us if the "foundation" of the heart (the Left Ventricular Function) is damaged, or is it just good for counting the beats?

How They Did It: The "Time-Travel" Comparison

The researchers looked at patients who had recently had a heart attack (a major event that damages the heart's "walls"). They gathered two types of data from these same patients:

The Hospital View: A standard 12-lead ECG taken while the patient was in the hospital.
The Home View: A 1-lead ECG recorded by the patient at home using a smartwatch (specifically a Withings watch).

They matched these recordings so they were looking at the same person, at roughly the same time. They then fed this data into a special type of computer brain (Deep Learning) that is designed to be "explainable."

The "Explainable" Part:
Usually, AI is a "black box"—it gives an answer, but you don't know why. This study used a special AI that acts like a holographic simulator. It can take a single heartbeat and say, "If we tweak this part of the wave to look like a healthy heart, here is what it would look like. If we tweak it to look like a damaged heart, here is that version." This lets doctors see exactly which parts of the signal the AI is using to make its decision.

The Results: The Single Camera is Surprisingly Good

The researchers trained the AI to predict if the heart's pumping function was normal, mildly damaged, moderately damaged, or severely damaged.

The Gold Standard: The 12-lead hospital ECG was the best, as expected. It got a score (AUC) of 0.897.
The Smartwatch: The single-lead smartwatch ECG got a score of 0.883.

The Takeaway: The smartwatch was almost as good as the full hospital machine! It captured nearly all the necessary information about the heart's structural health, even with just one "camera."

The Secret Sauce: Aggregation (The "Crowd Wisdom" Effect)

One interesting finding was about how the data was processed.

A hospital ECG is short (10 seconds).
A smartwatch recording is longer (30 seconds).

Because the smartwatch recording is longer, it contains many more heartbeats. The study found that if you let the AI look at all those heartbeats and take an average (like asking a crowd of people for an opinion rather than just one person), the smartwatch's accuracy went up significantly. It's like listening to a choir: one voice might be shaky, but the whole group singing together creates a clear, strong signal.

What the AI Actually "Saw"

When the researchers looked at the "explanation" part of the AI, they saw something fascinating:

The 12-lead model looked at many different angles (like checking the front, back, and sides of the house).
The 1-lead model focused on specific features in its single view, such as the height of the "R-peak" (the tall spike in the heartbeat) and the shape of the "T-wave" (the wave that follows the spike).

The study found that even with just one angle, these specific shapes changed in a way that clearly signaled if the heart muscle was damaged.

The Bottom Line

This paper proves that a simple, wearable smartwatch ECG isn't just for checking if your heart is skipping a beat. With the right AI, it can also give a very accurate estimate of how well your heart is pumping after a heart attack.

While it doesn't replace the full hospital test, it acts as a powerful, accessible "early warning system" that patients can use at home. The study also showed that by making the AI "explainable," we can understand why it thinks the heart is damaged, which helps build trust in these new tools.

Technical Summary: Leveraging Wearable 1-Lead ECGs for Cardiac Function Assessment

Problem Statement
Cardiovascular disease remains the leading cause of global mortality, necessitating accessible diagnostic tools. While the 12-lead electrocardiogram (ECG) is the clinical gold standard for assessing cardiac electrical activity and structural abnormalities, its utility is restricted to clinical settings due to equipment and personnel requirements. Conversely, wearable 1-lead ECGs (e.g., smartwatches) offer portability and enable remote monitoring but have historically been optimized primarily for rhythm analysis (such as atrial fibrillation detection) rather than structural or functional assessment. Although recent studies suggest 1-lead ECGs can detect left ventricular dysfunction (LVD), there is a lack of direct, head-to-head comparisons between wrist-based 1-lead ECGs and standard 12-lead ECGs acquired from the same patients, particularly within the post-myocardial infarction (MI) population. Furthermore, the application of explainable AI to interpret how single-lead signals correlate with left ventricular function (LVF) in this specific cohort remains underexplored.

Methodology
The study utilized a matched cohort of patients treated for acute coronary syndrome at Leiden University Medical Center (2018–2023). The dataset comprised:

1-Lead ECGs: 63,950 raw 30-second signals recorded at home using Withings ECG Smartwatches (300 Hz) from 708 patients.
12-Lead ECGs: 77,005 raw 10-second signals recorded in-hospital (500 Hz) from 6,703 patients.

Data Preprocessing and Labeling:
Signals were segmented into individual heartbeats (800 ms intervals centered on the R-peak). R-peak detection utilized a U-Net-based CNN. To address data quality issues inherent to home recordings (e.g., device orientation errors, non-resting states), an autocorrelation-based filter was applied to exclude low-quality 1-lead signals. LVF labels were derived from echocardiogram reports obtained within seven days of the ECG recording, classified into four ordinal categories (normal, mild, moderate, severe impairment) using a Large Language Model (LLM).

Model Architecture:
The study employed a partial $\beta$ -variational autoencoder (VAE) with seven recurrent convolutional blocks in both the encoder and decoder. The architecture was designed to separate latent representations into a non-linear signal component ( $z_{nls}$ ) and a linear signal component ( $z_{ls}$ ) directly linked to LVF prediction. Four distinct model configurations were trained and evaluated:

12-lead model: Trained on full 12-lead ECGs.
Lead I model: Trained on Lead I extracted from the 12-lead ECGs.
Pretrained 1-lead model: Pretrained on Lead I of 12-lead ECGs, then fine-tuned on wearable 1-lead ECGs.
1-lead model: Trained exclusively on wearable 1-lead ECGs.

For 1-lead inputs, the single channel was duplicated along the lead dimension to match the model's expected input shape ( $12 \times 400$ ).

Interpretability:
To ensure transparency, the study utilized attribute manipulation. By manipulating the $z_{ls}$ latent vector to represent hypothetical LVF states (from severe impairment to normal), the model reconstructed hypothetical ECG waveforms. This allowed for the visualization of specific morphological changes (e.g., R-peak amplitude, T-wave area) associated with different levels of LVF impairment.

Key Results

Predictive Performance: The full 12-lead model achieved the highest performance (AUC = 0.897). The wearable 1-lead model achieved an AUC of 0.863, while the pretrained 1-lead model achieved an AUC of 0.883. These results indicate that 1-lead ECGs can predict LVF with accuracy nearly matching that of 12-lead ECGs.
Sensitivity/Specificity Trade-offs: At a sensitivity of 0.90, the specificities were 0.64 for the 12-lead model, 0.71 for the pretrained 1-lead model, and 0.65 for the standard 1-lead model.
Aggregation Benefits: Heartbeat-level predictions were aggregated by averaging within each signal. This temporal aggregation significantly improved the performance of 1-lead models, suggesting that noise cancellation and the availability of more heartbeats in longer 30-second recordings compensate for reduced spatial resolution.
Consistency: All models demonstrated high heartbeat consistency scores (HBC > 0.95 for most), indicating robust intra-recording stability.
Interpretability Findings: Attribute manipulation revealed that the 12-lead model leveraged a diverse set of features across multiple leads (e.g., notched R-peaks in specific leads). In contrast, the 1-lead models relied on features resembling Lead I (e.g., R-peak amplitude, T-wave morphology), confirming that while single-lead models capture global abnormalities, they lack the spatial diversity of multi-lead systems.

Significance and Claims
The paper claims that wearable 1-lead ECGs, when combined with explainable deep learning, possess the capability to move beyond rhythm analysis to assess structural cardiac function, specifically left ventricular function in post-MI patients. The study demonstrates that:

Clinical Viability: 1-lead ECGs can provide clinically meaningful insights into LVF, achieving performance comparable to 12-lead ECGs, particularly when predictions are aggregated over time.
Explainability as an Enabler: The use of explainable AI (attribute manipulation) provides a transparent view of model reasoning, highlighting specific waveform components (like T-wave morphology) that drive predictions. This transparency is crucial for scaling these tools to resource-limited settings or home-based care where cardiology expertise may be unavailable.
Strategic Application: While 1-lead ECGs are not intended to replace 12-lead ECGs, they serve as a practical, scalable tool for screening and longitudinal monitoring, potentially acting as an early warning system to flag patients for further review.

The authors maintain a modest stance, acknowledging that 1-lead models exhibit a performance gap compared to 12-lead models and that their effectiveness may vary based on the location of myocardial infarction (e.g., global vs. regional abnormalities). They emphasize that future work is required to validate these findings in broader clinical subtypes and to refine data quality control for home-based acquisitions.