💻 computer science

From Reports to Ontologies: Ontology-Guided Representation Learning for 12-Lead ECG

The paper introduces MAR-ECG, an ontology-guided masked autoregressive framework that leverages a SNOMED-CT cardiac graph to supervise ECG representation learning without paired clinical reports, achieving state-of-the-art performance on downstream classification tasks by combining graph-smoothed contrastive learning with multi-scale physiological supervision.

Original authors: Lei Xu, Fahad Sohrab, Mehmet Yamac, Merja Heinaniemi, Moncef Gabbouj

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lei Xu, Fahad Sohrab, Mehmet Yamac, Merja Heinaniemi, Moncef Gabbouj

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 Picture: Teaching a Computer to "Read" a Heartbeat

Imagine a 12-lead ECG (electrocardiogram) as a complex, multi-channel symphony played by the heart. It's not just one line; it's 12 different "voices" recording the heart's electrical activity over time. Doctors look at these lines to diagnose problems, but teaching a computer to understand them is hard because the data is messy and the "meaning" is hidden in the patterns.

For a long time, researchers tried to teach computers to understand these heartbeats in two ways:

The "Guessing Game" (Signal-Only): They let the computer look at thousands of heartbeats and try to fill in missing pieces (like a puzzle). This helps the computer learn the shape of the waves, but it doesn't really teach the computer what those shapes mean medically.
The "Translator" (Text-Based): They paired heartbeats with the doctor's written notes (reports). The computer learns to match the squiggly lines to the words. The problem? These written reports are expensive to get, hard to translate between languages, and often just copy-paste templates that don't offer much new information.

The MAR-ECG Solution:
The authors propose a third way. Instead of using messy text reports, they use a structured map of medical knowledge called an "ontology" (specifically, a SNOMED-CT cardiac graph). Think of this as a highly organized family tree of heart diseases.

The Core Idea: The "Family Tree" Analogy

Imagine you are trying to teach a child the difference between a "Dog" and a "Cat."

The Old Way (Text): You show them a picture of a dog and read a long, complicated paragraph describing it.
The New Way (MAR-ECG): You show them a family tree. You say, "This is a 'Dog'. It is related to a 'Poodle' (its child) and a 'Wolf' (its cousin)."

In the MAR-ECG system, the "Family Tree" is a graph of 40 heart concepts (like "Heart Attack," "Irregular Rhythm," "Normal").

If the computer sees a heartbeat that looks like a "Heart Attack," it doesn't just learn that one label.
Because of the graph, it also learns that this is related to "Chest Pain" (a sibling concept) and "Ischemia" (a parent concept).
Crucially, it learns that "Heart Attack" is very different from "Normal Sinus Rhythm" (which are on opposite sides of the tree).

This allows the computer to learn the relationships between diseases without needing a human to write a paragraph about every single heartbeat.

How It Works: The Two-Step Dance

The MAR-ECG model uses two main tricks to learn, acting like a student studying for a test:

1. The "Soft" Lesson (Graph-Smoothed Contrastive Learning)

Usually, when a computer learns, it's told: "This is a Dog. If you think it's a Cat, you are 100% wrong." This is a "hard" lesson.
MAR-ECG uses a "soft" lesson based on the family tree.

The Analogy: If the computer thinks a "Heart Attack" is actually a "Normal Heartbeat," it gets a big penalty. But if it thinks a "Heart Attack" is actually "Chest Pain" (which is closely related), it gets a smaller penalty.
Why it helps: It teaches the computer that some mistakes are worse than others. It learns that "Anterior Heart Attack" and "Inferior Heart Attack" are cousins, so confusing them isn't as bad as confusing a Heart Attack with a Normal heartbeat. This "softening" helps the computer understand the nuances of the heart better.

2. The "Zoom-In" Lesson (Multi-Scale Physiological Supervision)

The first lesson looks at the whole heartbeat (the "rhythm"). But doctors also look at tiny details, like the shape of a specific bump in the wave (the QRS complex).

The Analogy: Imagine looking at a forest. The first lesson teaches you "This is a forest." The second lesson zooms in and says, "Look at this specific tree; it has 5 leaves and a broken branch."
How it works: The model automatically detects the "R-peaks" (the tallest spikes in the heartbeat) and checks if the tiny shapes around them make sense. It does this without needing a human to label every single spike. It just uses math to ensure the tiny details match the big picture.

The Results: Why It Matters

The researchers tested this new method on about 40,000 public heart recordings. They found that:

It beats the "Guessing Game": It learned much faster and better than methods that only looked at the signal shapes.
It rivals the "Translators": Even though it never saw a single written doctor's report, it performed just as well as (and sometimes better than) the most advanced models that did use text reports.
It's efficient: Because it uses the structured "Family Tree" instead of messy text, it doesn't need expensive licenses or translation services. It works with the data hospitals already have.

Summary

MAR-ECG is like a medical student who learns by studying a structured map of heart diseases rather than reading thousands of handwritten notes. By understanding how heart conditions are related to each other (like family members), the computer learns to recognize heart problems more accurately, even when it hasn't been explicitly told what every single heartbeat "means" in plain English.

Key Takeaway: You don't need a library of doctor's notes to teach a computer medicine; you just need a good map of how the diseases relate to one another.

Technical Summary: MAR-ECG

Problem Statement

The 12-lead electrocardiogram (ECG) is a complex spatiotemporal signal containing diagnostic information ranging from millisecond waveform morphology to multi-second rhythm dynamics. Existing representation learning approaches face two primary limitations:

Signal-only Self-Supervised Learning (SSL): Methods utilizing masked patch reconstruction, latent prediction, or contrastive learning on unlabelled data often learn low-level morphological structures but fail to capture the structured clinical relationships required for downstream diagnostic tasks.
Multimodal ECG-Text Alignment: While CLIP-style methods aligning ECG signals with paired clinical reports achieve state-of-the-art (SOTA) performance, they rely on expensive, licensed text corpora that do not transfer well across hospitals or languages. Furthermore, ECG reports are often templated, yielding pooled text embeddings that lack discriminative signal compared to free-text radiology reports.

Crucially, both paradigms overlook a structured resource available in virtually every clinical recording: SNOMED-CT diagnostic codes. These codes form a curated cardiac ontology that encodes clinical relationships (e.g., parent-child hierarchies, sibling co-occurrences) more precisely than free-text narratives and are standardized across institutions.

Methodology: MAR-ECG

The authors propose MAR-ECG (Masked Autoregressive Representation for ECG), an ontology-guided framework that supervises the encoder directly using a curated 40-node SNOMED-CT cardiac concept graph ( $G$ ) rather than paired clinical text. The framework combines a masked autoregressive (MAR) backbone with two complementary ontology-guided objectives.

1. Backbone Architecture

Input Tokenization: The 12-lead ECG signal is segmented into overlapping temporal patches per lead. A shared linear projection embeds these patches into tokens, augmented with learnable lead and position embeddings.
Encoder: A bidirectional transformer employs factorized spatial-then-temporal attention. It first attends across the 12 leads at a single timestep, then across time within a single lead. This reduces complexity from $O(C^2T^2)$ to $O(CT^2 + C^2T)$ while preserving the inductive bias that ECG patterns arise from interactions in both axes.
Masked Autoregressive Reconstruction: A fraction of (lead, patch) pairs are masked. A causal decoder predicts the original patch values conditioned on the encoder's full bidirectional context. This serves as the primary pretraining anchor.

2. Graph-Smoothed Contrastive Learning (GSCL)

This objective anchors the rhythm-pooled representation (aggregated across leads and time) to the SNOMED graph.

Soft Targets: Instead of hard one-hot labels, GSCL uses a soft target distribution derived from the graph-geodesic distance ( $D_{tree}$ ) between the primary positive concept ( $c^*_b$ ) and all other nodes in the graph.
Mechanism: The target probability for a concept $c$ is proportional to $\exp(-D_{tree}[c^*_b, c] / \sigma)$ . This ensures that clinically related concepts (e.g., parent roots, sibling shortcuts) act as "soft positives" rather than hard negatives, reinforcing the encoder's understanding of clinical hierarchy.
Prototype Generation: Concept prototypes are dynamically generated via a lightweight Graph Convolutional Network (GCN) applied to the fixed graph adjacency, ensuring structural similarity is injected into the representation space.

3. Multi-Scale Physiological Supervision (MSPS)

To address gaps in GSCL (which operates at the sample level), MSPS introduces lightweight auxiliary heads on the per-patch (lead-mean) representation:

Patch Rhythm Auxiliary: Predicts sample-level rhythm statistics (mean R-R interval, variability, heart rate bucket, alternation flag) broadcast to all patches.
Patch Position Auxiliary: Predicts sequence position buckets and R-peak phases (PRE-R, R, ST, T-WAVE) based on unsupervised R-peak detection.
Benefit: This extends supervision to the beat and rhythm scales without requiring additional annotations, preventing token collapse and anchoring the model to physiological landmarks.

4. Training Objective

The total loss combines the masked autoregressive reconstruction loss with GSCL, MSPS, and standard auxiliary regularizers (latent prediction, view contrast, and a legacy multi-prototype concept-text alignment head). The decoder is discarded after pretraining; only the encoder is retained for downstream evaluation.

Key Contributions

Ontology-Guided Pretraining: MAR-ECG is the first self-supervised ECG method to use an ontology-graph structure as the primary pretraining objective, eliminating the need for paired clinical text.
Graph-Smoothed Contrastive Learning: It introduces a novel contrastive objective that softens the InfoNCE target distribution using graph-geodesic distances, allowing clinically related concepts to reinforce one another.
Multi-Scale Physiological Supervision: It extends supervision beyond the sample level to the patch and beat levels using automatically extracted physiological statistics, bridging the gap between morphology and rhythm dynamics.
Resource Efficiency: By leveraging SNOMED-CT codes (inherent in clinical workflows) instead of free-text reports, the method avoids licensing barriers and cross-language transfer issues.

Results

The model was pretrained on $\sim$ 40,000 publicly available 12-lead ECGs with SNOMED-CT codes and evaluated via frozen linear probing on five downstream classification benchmarks (PTB-XL and CPSC2018 subsets).

Performance vs. Baselines: MAR-ECG consistently outperforms strong masked-autoregressive baselines (e.g., C1 configuration). In the low-label regime (1% and 10% labels), adding GSCL (C2) yields significant gains (e.g., +16.99 AUC on PTBXL-Rhythm at 1% labels).
Comparison to Multimodal Methods: Despite the absence of paired clinical text, MAR-ECG achieves performance competitive with SOTA multimodal ECG-text methods (e.g., MERL, MELP, D-BETA). Specifically, MAR-ECG (C2) matches or surpasses multimodal baselines on 9 of 15 evaluation cells.
Ablation: The addition of Multi-Scale Physiological Supervision (C3) further sharpens performance on rhythm-heavy tasks, though it incurs a minor trade-off on morphology-only tasks.

Significance and Claims

The paper claims that structured diagnostic ontologies can effectively substitute for free-text supervision in ECG representation learning. By aligning the encoder directly with the SNOMED-CT graph, MAR-ECG captures the clinical relationships necessary for downstream tasks without the data constraints of multimodal approaches. The authors posit that this approach offers a scalable, license-free alternative for developing robust ECG foundation models, particularly in settings where paired clinical text is unavailable or non-standardized. The implementation is made available to facilitate reproducibility.