⚡ electrical engineering

Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations

The paper introduces LuminaECG, a clinically structured framework that enhances ECG diagnosis by rendering signals as grid-based visual primitives with explicit wave boundaries, enabling a lightweight vision-language model to achieve expert-level reasoning and cross-population generalization by aligning measurable waveform evidence with doctor-grounded priors.

Original authors: Hongxiang Gao, He-yang Xu, Yuwen Li, Minghui Zhao, Zhipeng Cai, Xingyao Wang, Chenxi Yang, Jianqing Li, Chengyu Liu

Published 2026-08-11

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Hongxiang Gao, He-yang Xu, Yuwen Li, Minghui Zhao, Zhipeng Cai, Xingyao Wang, Chenxi Yang, Jianqing Li, Chengyu Liu

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 trying to teach a robot how to read a map. You could show it millions of pictures of maps and hope it figures out the rules on its own, or you could sit it down and say, "See this blue line? That's a river. See this squiggly black line? That's a road. Measure the distance between them with a ruler." This paper lives in the world of medical artificial intelligence, specifically looking at how computers interpret electrocardiograms, or ECGs. An ECG is a squiggly line drawing of your heart's electrical activity, printed on special grid paper where every tiny square represents a specific amount of time and voltage. Doctors don't just glance at these lines and guess; they act like detectives, measuring the distance between the waves with a ruler, counting the squares, and using those exact numbers to decide if a heart is healthy or sick. The big question researchers have been asking is: Can we just build a super-smart, giant computer brain and let it figure out the rules by itself, or do we need to teach the computer to read the map the same way a human doctor does, step-by-step with a ruler?

The researchers behind this study, who call their creation LUMINAECG, decided to try the "teach the rules" approach. They built a system that doesn't just look at the raw squiggly lines; instead, it first draws the ECG on a perfect digital grid, just like a doctor would see it on paper. Then, it highlights the different parts of the heartbeat in different colors—green for the P-wave, red for the QRS complex, and blue for the T-wave—making it impossible to miss the key features. They then trained a relatively small computer model (only 2 billion parameters, which is tiny compared to the massive models everyone else is using) to read these colorful, grid-lined images and write a report that includes the actual measurements, like "the heart rate is 60 beats per minute" or "the gap between waves is 180 milliseconds."

The results suggest that this "doctor-style" training works incredibly well. When they tested LUMINAECG, it didn't just sound like a doctor; it actually measured like one. On a standard test, it guessed the heart rate with an error of only 0.43 beats per minute, whereas the biggest, most powerful AI models made mistakes that were fifteen times larger. It also became much better at spotting dangerous heart problems that the giant models completely missed. For instance, on a test involving urgent heart conditions, LUMINAECG found about half of the critical issues, while the massive models found almost none. Perhaps most surprisingly, even though the model was never taught to predict who might get sick in the future, the reports it wrote contained hidden clues that helped predict patient outcomes better than just knowing their age and gender.

The paper argues that the secret to making medical AI better isn't just making the computer brain bigger and more expensive. Instead, it suggests that the real breakthrough comes from how we teach the computer. By forcing the AI to use the same "ruler and grid" method that human doctors have used for decades, a small, efficient model can learn to reason like a human expert. The authors found that without this structured, measurement-based training, even the most powerful AI models tend to guess or hallucinate, missing the tiny details that save lives. So, the next time you hear about a new medical AI, the paper suggests we shouldn't just ask, "How big is it?" but rather, "Does it know how to use a ruler?"

Technical Summary: Diagnosing as Cardiologists Do (LUMINAECG)

Problem Statement

Current automated Electrocardiogram (ECG) interpretation systems face a fundamental disconnect between model output and clinical reasoning. While deep learning models have achieved high accuracy in closed-set classification tasks, they often function as "black boxes" that emit diagnostic labels without providing the localized, quantitative evidence required for clinical verification. Furthermore, emerging Multimodal Large Language Models (MLLMs) and vision-language models, while capable of generating fluent text, frequently lack the domain-specific priors necessary for precise waveform localization, interval measurement, and the explicit linking of morphological evidence to diagnostic criteria. Consequently, these models struggle to generalize across diverse populations and devices, often relying on dataset-specific shortcuts rather than transferable clinical knowledge. The central question addressed is whether the structured, measurement-grounded reading process used by human cardiologists can serve as an effective prior to elevate ECG agents beyond simple pattern recognition toward clinically meaningful reasoning.

Methodology: LUMINAECG

The authors introduce LUMINAECG, a clinically structured ECG reasoning framework that reformulates ECG interpretation as a "measurement-grounded visual reading" task. Rather than scaling up model architecture, the approach focuses on embedding expert diagnostic priors directly into the training data construction.

1. Data Construction Pipeline

The training data is derived from the MIMIC-IV-ECG dataset (800,035 twelve-lead recordings). The pipeline processes each record through three offline stages to create a "doctor-aligned" supervision signal:

Measurement Extraction: A NeuroKit-based delineator identifies fiducial points (P, Q, R, S, T, U) across all 12 leads, computing a 303-field clinical feature panel. This includes wave amplitudes, ST levels, intervals, axes, and rule-based rhythm/conduction scores. Crucially, all numerical values in the subsequent report are derived strictly from these measurements, not free-text labels.
Image Rendering: ECG signals are rendered onto standard electrocardiographic grid paper (25 mm/s, 10 mm/mV) to preserve spatial and scale cues. A color-coded segmentation overlay is applied: P-waves (green), QRS complexes (red), and T-waves (blue). This makes waveform components explicitly identifiable to the visual encoder.
Grounded Report Generation: An LLM (GPT-4o) generates structured, eight-section reports (Rhythm, Intervals, Axis, Conduction, ST-T, Signal Quality, Evidence, Conclusion) based only on the extracted measurements and device diagnoses. A grounding protocol ensures the report does not hallucinate findings absent from the measurements and explicitly links every conclusion to specific evidence (e.g., "First-degree AV block" anchored to "PR > 200 ms").

2. Model Architecture

LUMINAECG utilizes a compact 2-billion parameter (2B) vision-language backbone (based on Qwen2B). The model is fine-tuned using Low-Rank Adaptation (LoRA) without any architectural modifications or additional alignment procedures. The training objective is to map the wave-annotated ECG images and fixed instructions to the grounded, structured reports. This design isolates the impact of clinically informed data construction from model scaling.

Key Contributions

Clinically Structured Supervision: The paper proposes that embedding the "grammar" of clinical reading (localization, measurement, and evidence-to-diagnosis mapping) into the training data is more effective than simply increasing model size.
Measurement-Grounded Interpretation: Unlike prior works that generate text descriptions, LUMINAECG is trained to produce reports where diagnostic claims are inextricably linked to quantitative waveform measurements and visual segmentation cues.
Compact Model Performance: The study demonstrates that a 2B parameter model, when trained with doctor-grounded priors, can outperform significantly larger zero-shot models (including 235B parameter models) and specialized ECG classifiers on critical clinical tasks.

Results

The framework was evaluated against open, proprietary, and ECG-specialist zero-shot baselines across multiple benchmarks:

1. Report Generation and Measurement Accuracy

On the held-out MIMIC-IV-ECG test set ( $n=79,981$ ):

Report Quality: LUMINAECG achieved superior text-similarity scores (BLEU-4: 0.741, CIDEr: 0.896) compared to frontier models like GPT-5.4 (BLEU-4: 0.307) and Claude-Opus-4.7 (BLEU-4: 0.292).
Quantitative Fidelity: The model achieved a Heart Rate Mean Absolute Error (MAE) of 0.43 bpm, approximately 15 times lower than the best zero-shot baseline (6.45 bpm). It also significantly outperformed baselines in measuring PR intervals (8.97 ms MAE) and QRS duration (4.48 ms MAE).

2. Critical Finding Recovery

On the PTB-XL dataset ( $n=4,396$ ), LUMINAECG recovered 50.8% of critical findings (e.g., atrial fibrillation, ST elevation, LV hypertrophy). In contrast, all evaluated frontier zero-shot systems remained below 10.3%, and the specialized ECG-R1 model recovered only 1.1%. Notably, LUMINAECG maintained a balanced error profile (low hallucination and low omission), whereas frontier models exhibited either severe hallucination or omission.

3. Human-Reader Alignment (CODE-test)

On the CODE-test benchmark ( $n=827$ ), LUMINAECG achieved a macro-F1 of 0.840 and micro-F1 of 0.852. This performance places the model:

Above the medical-student tier (0.817) and emergency-resident tier (0.829).
Approaching the cardiology-resident tier (0.862).
It surpassed the medical-student tier on five of the six red-flag diagnoses.

4. Cross-Cohort Generalization

Without retraining, LUMINAECG transferred effectively to external datasets from Brazil (CODE-test), China (Chapman), and Germany (PTB-XL), maintaining macro-F1 scores between 0.644 and 0.840. In comparison, a ResNet1d-101 classifier trained on the same source data degraded sharply, dropping to a macro-F1 of 0.049 on PTB-XL.

5. Emergent Prognostic Value

In an external cohort (Sami-Trop, $n=1,631$ ) with no training contact, features extracted from LUMINAECG's structured reports improved the C-index of an age-and-sex Cox model from 0.678 to 0.764 ( $\Delta C = +0.086, P=0.005$ ). This indicates the model learned outcome-relevant structures despite never being trained on mortality data.

Significance and Claims

The paper argues that the bottleneck in medical multimodal modeling lies not in model capacity, but in the alignment between measurable waveform evidence and clinical knowledge. The authors claim that:

Supervision Design > Model Scale: Effective ECG agents require supervision that preserves how clinicians localize evidence, derive measurements, and connect them to diagnoses. A compact 2B model with this supervision outperforms massive zero-shot models.
Structured Reasoning over Imitation: Doctor-grounded supervision transforms ECG reporting from superficial language imitation into measurement-grounded interpretation, ensuring that diagnostic claims are anchored in verifiable evidence.
Transferable Clinical Knowledge: By learning to organize evidence through report-like priors, the model acquires transferable diagnostic structures that generalize across geographically and demographically diverse cohorts, overcoming the dataset-specific correlations that plague conventional classifiers.
Clinical Readiness: The model's ability to reach the performance tier of junior human readers and generate reports with emergent prognostic signals suggests that expert-informed data design is a fundamental prerequisite for deploying reliable medical AI, potentially more so than architectural scaling.

The authors conclude that future medical agent development should prioritize clinically structured supervision and human-readable evidence over purely architectural expansion.