🤖 AI

ECG-LENS: Lead-Aware Clinical Context Enriched ECG Report Generation and Evaluation

The paper introduces ECG-LENS, an end-to-end framework that integrates multi-lead signal modeling with clinically enriched text generation to produce high-quality ECG reports, outperforming state-of-the-art methods on PTB-XL and MIMIC-IV-ECG datasets while also proposing a novel ECG-specific evaluation metric called F1-ECGBERT.

Original authors: Akanta Das, Tasinul Islam Ahon, Ahmed Mahir Sultan Rumi, Md Mahbubur Rahman, Tausif Amim Shadly, Tanzima Hashem

Published 2026-08-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Akanta Das, Tasinul Islam Ahon, Ahmed Mahir Sultan Rumi, Md Mahbubur Rahman, Tausif Amim Shadly, Tanzima Hashem

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

Imagine you are a detective trying to solve a mystery, but instead of looking at fingerprints or footprints, you are listening to the electrical whispers of a heart. This is the world of electrocardiography, or ECG. For decades, doctors have used these squiggly lines to spot trouble, but reading them is like trying to understand a complex symphony just by looking at the sheet music; it requires years of training to hear the difference between a harmless skip and a dangerous rhythm. Recently, scientists have been teaching computers to do this listening for us, hoping to give a second opinion to doctors everywhere. But here's the catch: while computers are getting good at saying "Yes, there is a problem," they are still terrible at writing the actual report that explains what the problem is, where it is, and why it matters. It's like having a robot that can spot a fire but only says "Fire!" without telling you which room is burning or if you need a hose or a bucket.

This is where a new team of researchers steps in with a clever solution called ECG-LENS. Think of the heart's electrical signal as a choir of twelve different singers (the twelve leads of an ECG). Old computer models often forced all twelve singers to blend into one giant, muddy voice before trying to understand the song, which meant they missed the unique notes each singer was hitting. ECG-LENS, however, acts like a super-smart conductor who listens to every single singer individually and how they harmonize together. It doesn't just guess the diagnosis; it uses a "reference" of likely conditions to guide a language writer (a type of AI) to compose a report that sounds like a real doctor wrote it. The result? A system that doesn't just find the error but writes a clear, accurate story about the heart's condition, even when tested on data it has never seen before.

The Problem: The "Muddy Voice" of Old AI

For a long time, artificial intelligence in medicine has been great at sorting things into boxes. If you show an AI an ECG, it can usually tell you if the heart is normal or if there's a specific disease. But doctors don't just need a box labeled "Abnormal"; they need a paragraph that says, "The heart is beating in a chaotic rhythm, and the lower part of the heart shows signs of an old injury."

Previous attempts to make computers write these reports had a major flaw. They treated the twelve different electrical views of the heart as one big, jumbled mess. Imagine trying to describe a painting by squinting at it until all the colors blur into a single shade of brown. You might know it's a painting, but you can't tell if it's a sunset or a storm. In the same way, early AI models mashed the twelve leads together too early, losing the specific clues that tell a doctor exactly where the trouble is. They also struggled with the language, often producing reports that sounded fluent but missed the critical medical details, or worse, made up facts that weren't there.

The Solution: ECG-LENS and the "Smart Conductor"

The researchers behind ECG-LENS decided to fix this by changing how the computer "listens" to the heart. They built a system with two main listening ears:

The Individual Ears (Lead-Aware Encoders): Instead of blending the twelve leads, the system gives each lead its own dedicated microphone. It uses a specialized tool (a ResNet-18 encoder) to listen to the unique shape and rhythm of each lead separately. This ensures that if one singer (lead) is off-key, the computer knows exactly which one it is, rather than assuming the whole choir is just "a little off."
The Big Picture Ear (Global Encoder): While listening to the individuals, the system also has a second ear that listens to the whole choir at once. This helps it understand how the leads work together, spotting patterns that only appear when you see the full picture.

But listening isn't enough; the computer needs to know what to say. This is where the "reference" comes in. Before writing the report, the system runs the ECG through a separate, highly skilled diagnostic tool (called MERL) to get a list of the most likely conditions. It turns these medical labels into a simple, clear sentence—a "clinical prompt." It's like giving the writer a headline: "The heart shows signs of an old injury in the lower section and a chaotic rhythm." The writing AI then uses this headline, combined with the detailed audio from the twelve leads, to craft the final report.

To make sure the training data was perfect, the team also cleaned up the "textbooks" they used. Real medical reports are often messy, filled with repetitive notes or administrative jargon. The team used a powerful AI to strip away the fluff, leaving only the pure, clinically important facts. This taught the new model to focus on what actually matters, ignoring the noise.

The Results: A Clearer Voice

When the team tested ECG-LENS, the results were impressive. They compared it to the best existing methods using standard tests for writing quality and, more importantly, a new test they invented called F1-ECGBERT. This new test doesn't just check if the words match; it checks if the medical facts match. It asks, "Did the computer correctly identify the same diseases as the human doctor?"

On the PTB-XL dataset (a large collection of heart recordings), ECG-LENS beat the competition by a wide margin. It improved the quality of the writing by 4.0% in one measure, 6.3% in another, and a massive 11.5% in the medical fact-checking score. Even more exciting, when they tested it on a completely different set of data from the US (MIMIC-IV-ECG) that it had never seen before, it still performed better than any other method. This suggests the system isn't just memorizing answers; it's actually learning how to understand the heart.

A trained cardiologist reviewed a random selection of the reports and found that 67% were completely correct, with nearly all the rest being partially correct. The system was also incredibly fast, taking only about 30 milliseconds to generate a report, which means it could be used in real-time, even in places with limited computer power.

What It Means

The paper suggests that by treating the heart's electrical signals with more respect—listening to each lead individually and giving the writer a clear plan to follow—we can create AI that is not just a calculator, but a helpful partner. While the system isn't perfect (it still makes mistakes in rare cases and can occasionally "hallucinate" details), it represents a significant step forward. It moves us away from simple "yes/no" diagnoses toward rich, accurate stories that could help doctors everywhere, especially in rural areas where specialists are hard to find, to understand the heart's story with greater clarity and speed.

Technical Summary: ECG-LENS

Problem Statement

Electrocardiography (ECG) is a primary non-invasive tool for diagnosing cardiovascular diseases, yet transforming multi-lead recordings into reliable clinical reports remains a significant challenge. While deep learning has advanced ECG classification, automated report generation lags behind image-based domains like radiology. Existing ECG report generation systems suffer from three primary limitations:

Loss of Lead-Specific Evidence: Most methods compress the entire multi-lead recording into a single global representation early in the pipeline, diluting localized morphological patterns (e.g., ST elevation specific to leads II, III, and aVF) that are critical for diagnosis.
Weak Clinical Grounding: Current approaches often condition language models on global signal embeddings without an explicit diagnostic plan, leading to reports that lack coherent clinical reasoning or contain hallucinations.
Inadequate Evaluation: Traditional lexical metrics (BLEU, ROUGE, METEOR) measure surface-level text similarity but fail to assess whether the generated report preserves clinically correct diagnostic information or distinguishes valid paraphrasing from clinically incorrect hallucinations.

Methodology: ECG-LENS

ECG-LENS (Lead-aware Enrichment with Context Narrative Synthesis) is an end-to-end framework designed to jointly model multi-lead signals, diagnosis-aware representations, and clinically grounded report generation. The architecture consists of four integrated components:

1. Lead-Aware Encoding Module

Instead of flattening all leads into a single sequence, ECG-LENS employs a dual-pathway encoding strategy:

Lead-Specific Encoders: Each of the 12 leads is processed independently using a 1D ResNet-18 encoder. This preserves localized waveform morphology and lead identity, ensuring that patterns specific to certain lead groups (e.g., inferior vs. anterior leads) are not obscured.
Global Context Encoder: A separate global 1D ResNet-18 encoder processes the stacked multi-channel signal ( $L \times T$ ) to capture inter-lead dependencies and global cardiac context.
Fusion: The lead-specific features and the global feature are concatenated to form a comprehensive representation ( $H_{ECG}$ ) that retains both local details and global relationships.

2. Clinically Enriched Prompt Generation

To guide the decoder with explicit diagnostic context, the framework utilizes a frozen, pre-trained ECG classifier (MERL) to predict diagnostic labels.

Zero-Shot Classification: MERL provides confidence scores for four diagnostic categories: superclass, subclass, rhythm, and form.
Prompt Construction: High-confidence predictions are mapped to standardized clinical phrases (following SCP-ECG standards) and organized into a coherent paragraph. This prompt serves as an intermediate clinical representation, providing the decoder with a structured diagnostic plan before generating the final text.

3. Clinically Focused Report Preprocessing

Raw ECG reports often contain repetitive, non-informative, or multilingual text that introduces noise. The authors employ a few-shot prompting strategy with GPT-5.5 to refine reference reports. This process:

Removes administrative text, isolated voltage measurements, and redundant statements.
Standardizes language and translates multilingual reports into consistent clinical English.
Ensures that only clinically meaningful findings (rhythm, morphology, diagnosis, lead associations) are preserved as training targets.

4. Report Generation

A lightweight GPT-2 style autoregressive decoder (trained from scratch with a domain-specific tokenizer) generates the final report. The generation is conditioned on:

The combined local-global ECG representation ( $H_{ECG}$ ).
The clinically enriched diagnostic prompt.
The preceding generated tokens.

The model is optimized end-to-end to minimize the negative log-likelihood of the target reports, with gradients flowing through the encoders, projection layers, and decoder, while the MERL classifier remains frozen.

Key Contributions

Lead-Aware Representation Learning: A novel architecture combining lead-wise ResNet encoders with a global encoder to preserve localized morphology while modeling inter-lead dependencies.
Clinical Context-Guided Generation: A mechanism that uses a frozen classifier to generate structured clinical prompts, providing the decoder with an explicit diagnostic plan to improve coherence and grounding.
Report Preprocessing Strategy: A refinement pipeline that standardizes multilingual and abbreviated reports, reducing ambiguity and focusing training on clinically significant content.
F1-ECGBERT Metric: An ECG-specific evaluation metric based on BERT classifiers that measures diagnostic agreement between generated and reference reports, addressing the limitations of lexical overlap metrics.
Comprehensive Evaluation: Demonstrated performance improvements on both in-domain (PTB-XL) and cross-domain (MIMIC-IV-ECG) datasets.

Experimental Results

Experiments were conducted on the PTB-XL dataset for training and in-domain testing, with cross-domain evaluation on MIMIC-IV-ECG.

Textual Quality: ECG-LENS outperformed state-of-the-art baselines (including MEIT, BiECG-LLM, and HeartLLM) across all standard NLP metrics. On PTB-XL, it achieved absolute gains of 4.0% in METEOR, 6.3% in ROUGE-L, and 11.5% in F1-ECGBERT over the strongest baselines.
Diagnostic Consistency: Using the new F1-ECGBERT metric, ECG-LENS achieved the highest scores across all four diagnostic categories (superclass, subclass, rhythm, form), exceeding the best baseline by 10.9% in the superclass category.
Cross-Domain Generalization: The model maintained superior performance on MIMIC-IV-ECG, showing significant gains in BLEU scores (e.g., +12.8% in BLEU-2), indicating robust learning of ECG-report relationships rather than dataset-specific wording.
Ablation Studies: The study confirmed that the combination of lead-specific encoders, report preprocessing, and clinical prompting yields cumulative improvements. The ResNet-based global encoder consistently outperformed Transformer-based alternatives in this context.
Clinical Review: A trained cardiologist reviewed a subset of generated reports, judging 67% as fully clinically correct and nearly all remaining reports as partially correct.

Significance and Claims

The paper claims that ECG-LENS represents a significant step toward reliable, automated ECG report generation by addressing the specific complexities of multi-lead time-series data. By integrating lead-aware encoding with clinically enriched prompting, the framework produces reports that are not only linguistically fluent but also diagnostically consistent. The introduction of F1-ECGBERT provides a more clinically meaningful evaluation standard than traditional lexical metrics. The authors note that the model is lightweight with an inference time of approximately 30ms, making it suitable for resource-constrained settings. However, they modestly acknowledge limitations, including the potential for hallucination in rare cases and the computational constraints that prevented testing larger-capacity architectures.