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

Summary

This paper introduces LUMINA ECG, a clinically structured ECG reasoning framework that reformulates ECG interpretation as measurement-grounded visual reading, guided by the hypothesis that effective ECG agents require not only larger models, but supervision that preserves the alignment between measurable waveform evidence and clinical knowledge.

Core Approach: The framework encodes cardiologist diagnostic priors into the data construction process. ECG signals are rendered on standard electrocardiographic grid paper (25 mm/s, 10 mm/mV) to preserve spatial and scale cues. P-wave, QRS-complex, and T-wave boundaries are explicitly delineated with color-coded segmentation (P=green, QRS=red, T=blue), decomposing the waveform into discrete visual measurement primitives. A general 2B vision–language backbone (Qwen2B) is then trained with low-rank supervised fine-tuning (LoRA) without architectural modification. Training data comes from MIMIC-IV-ECG (800,035 recordings, split 9:1 into 719,828 training and 79,981 held-out records), with each record passing through three stages: measurement extraction (a 303-field clinical feature panel), wave-annotated image rendering, and grounded report generation (eight-section Chinese reports generated by gpt-4o from measurements only, with a grounding protocol constraining assertions to device diagnoses and measurements).

Key Results:

1. Report Generation and Measurement Accuracy: On the MIMIC-IV-ECG held-out test set (n=79,981), LUMINA ECG achieved BLEU-1 0.853, BLEU-4 0.741, ROUGE-L 0.808, and CIDEr 0.896, substantially outperforming frontier zero-shot systems (GPT-5.4: BLEU-4 0.307, ROUGE-L 0.392; Claude-Opus-4.7: BLEU-4 0.292, ROUGE-L 0.383). For numeric measurements, LUMINA ECG achieved heart-rate MAE of only 0.43 bpm (versus 6.45 bpm for GPT-5.4, 8.37 bpm for Claude-Opus-4.7, 15.73 bpm for Qwen3-VL-8B), PR MAE 8.97 ms, and QRS MAE 4.48 ms (versus 24–26 ms and 14–21 ms for the strongest zero-shot systems).

2. Critical-Finding Recovery: On PTB-XL (n=4,396, no training contact), across 2,569 critical-finding instances spanning eleven urgent diagnoses, LUMINA ECG-2B recovered 50.8%, while all frontier zero-shot systems remained below 10.3% (Claude-Opus-4.7: 0.096, GPT-5.4: 0.076, Gemini-3.1-Pro: 0.051, Qwen3-VL-235B-A22B: 0.008). The paper notes A 2-billion-parameter doctor-grounded model thus exceeded a 235-billion-parameter zero-shot model on this endpoint by roughly 64×. The ECG-specialist ECG-R1 recovered only 1.1% of critical findings on in-domain MIMIC-IV-ECG. LUMINA ECG was strongest on eight of eleven critical findings (atrial fibrillation recall 0.92, LV hypertrophy 0.78, LBBB 0.70). Frontier models showed either loud hallucinators (63–72% hallucinated claims) or a silent omitter (Gemini, 64% omissions), while LUMINA ECG occupied a balanced low-error regime (hallucination 0.21, omission 0.21).

3. Human-Reader Comparison on CODE-test: On CODE-test (n=827), LUMINA ECG-2B reached macro-F1 0.840 and micro-F1 0.852 over six red-flag diagnoses, exceeding the medical-student tier (0.817/0.833) and emergency-resident tier (0.829/0.846), and approaching the cardiology-resident tier (0.862/0.876). It exceeded medical-student performance on five of six diagnoses (first-degree AV block 0.818 vs 0.732, LBBB 0.929 vs 0.915, sinus bradycardia 0.778 vs 0.750, atrial fibrillation 0.765 vs 0.706, sinus tachycardia 0.909 vs 0.873), with RBBB remaining below both junior tiers (0.841 vs 0.928 and 0.852).

4. Cross-Cohort Generalization: Across three external cohorts with no training contact, LUMINA ECG-2B retained macro-F1 between 0.644 and 0.840 (CODE-test Brazil: 0.840, Chapman–Shaoxing China: 0.644, PTB-XL Germany: 0.678). In contrast, a ResNet1d-101 waveform classifier trained on the same MIMIC source data dropped to 0.470 on CODE-test, 0.350 on Chapman–Shaoxing, and 0.049 on PTB-XL, representing gains of +0.370, +0.294, and +0.629 macro-F1 for LUMINA ECG. Atrial fibrillation remained strong across Chapman and PTB-XL (F1 0.898 and 0.892), and sinus tachycardia transferred consistently (F1 0.889 on Chapman, 0.836 on PTB-XL).

5. Emergent Prognostic Information: On the external Sami-Trop cohort (Chagas cardiomyopathy; n=1,631, 104 deaths, median follow-up 2.1 years, no training contact), features extracted from LUMINA ECG's reports added prognostic value over age and sex. The concordance index of a demographic Cox model rose from 0.678 to 0.764 (∆C = +0.086, 95% CI +0.026 to +0.147, P = 0.005), with integrated discrimination improvement +0.047 and category-free net reclassification index +0.36. The risk score was well calibrated (slope 1.18, r = 0.98) and stratified two-year mortality from 0.6% in the lowest decile to 22.3% in the highest. Significant age- and sex-adjusted hazard ratios were found for major-arrhythmia flag (2.88, 95% CI 1.86–4.46) and prolonged QTc (3.94, 2.49–6.24).

Conclusions: The paper argues that "the bottleneck in medical multimodal modelling lies not only in model capacity, but also in whether the training data preserve how clinicians localize evidence, derive measurements, and connect those measurements to diagnostic conclusions. It suggests that expert diagnostic priors should be treated as a foundation for medical multimodal models rather than expected to emerge from scale alone. Future work should incorporate lead-level evidence, morphology-specific descriptions, physician-verified rationales, and multimodal clinical context to move from report-like generation to explicit clinical reasoning."

Improvements to AI Systems Based on LUMINA ECG

1. Measurement-Grounded Visual Supervision for Medical Imaging Models

• Encode clinical priors (e.g., waveform segmentation, grid-scale rendering) directly into training data, not just model architecture.

• Use color-coded anatomical/physiological primitives (P, QRS, T) to force the model to localize evidence before generating text.

• Resulting capability: AI can produce reports with near-zero numeric error (heart-rate MAE 0.43 bpm vs. 6–16 bpm for zero-shot LLMs) and reduce hallucination/omission rates to 21% each, versus 63–72% hallucination in frontier models.

2. Low-Rank Fine-Tuning of Small Vision–Language Backbones with Domain-Specific Data

• Apply LoRA to a 2B-parameter backbone (e.g., Qwen2B) using a large, structured clinical dataset (800k+ ECG records) rather than relying on larger models.

• Resulting capability: A 2B model outperforms a 235B zero-shot model by 64× on critical-finding recovery (50.8% vs. 0.8%), while remaining computationally efficient and deployable in resource-constrained clinical settings.

3. Grounded Report Generation Protocol with Constrained Assertions

• Train models to generate reports only from device-derived measurements and explicit diagnostic labels, with a protocol that forbids unsupported claims.

• Resulting capability: AI systems can produce clinically safe, evidence-aligned reports that match or exceed human junior doctors (macro-F1 0.840 vs. medical-student 0.817 and emergency-resident 0.829 on CODE-test), while avoiding both loud hallucination and silent omission failure modes.

4. Cross-Cohort Generalization via Visual Measurement Primitives

• Replace raw waveform classification with visual grid-paper rendering and explicit segmentation, which preserves scale and spatial cues that generalize across devices, populations, and recording formats.

• Resulting capability: AI retains high diagnostic accuracy across diverse external cohorts (macro-F1 0.644–0.840) and improves over waveform-only classifiers by +0.294 to +0.629 macro-F1, enabling deployment in new hospitals without retraining.

5. Emergent Prognostic Feature Extraction from Structured Reports

• Use the model’s generated reports as a feature source for downstream survival analysis, not just classification.

• Resulting capability: AI can extract clinically meaningful prognostic markers (e.g., prolonged QTc hazard ratio 3.94, major-arrhythmia flag 2.88) that improve mortality prediction over demographics alone (C-index +0.086, P=0.005), enabling risk stratification for diseases like Chagas cardiomyopathy.

6. Balanced Error Regime for Critical-Finding Triage

• Optimize for joint minimization of hallucination and omission, rather than maximizing raw accuracy.

• Resulting capability: AI can serve as a reliable triage tool for urgent ECG findings (e.g., atrial fibrillation recall 0.92, LV hypertrophy 0.78, LBBB 0.70), recovering 50.8% of critical instances where all frontier zero-shot systems remain below 10.3%, thus reducing missed diagnoses in emergency settings.

7. Data-Centric Construction Pipeline for Medical Multimodal Training

• Adopt a three-stage pipeline (measurement extraction → annotated image rendering → grounded report generation) that can be replicated for other signal-based modalities (e.g., EEG, EMG, continuous glucose monitoring).

• Resulting capability: AI systems can be built for other medical signals with similar measurement-grounded supervision, achieving high accuracy and generalization without requiring massive proprietary models or manual annotation.

Cardiologists interpret electrocardiograms by localizing waveform components, measuring rhythm and interval patterns, and translating these structured observations into diagnostic evidence. Whether this expert reading process can serve as an effective prior for ECG agents remains unclear. To address this question, we introduce LuminaECG, a clinically structured ECG reasoning framework that reformulates ECG interpretation as measurement-grounded visual reading. ECG signals are rendered on standard electrocardiographic grid paper to preserve the spatial and scale cues used in clinical reading. P-wave, QRS-complex, and T-wave boundaries are explicitly delineated, and color-coded segmentation decomposes the waveform into discrete visual measurement primitives. A general 2B vision-language backbone is then trained with low-rank supervised fine-tuning to associate these primitives with diagnostic reasoning, without architectural modification. Across open, proprietary, and ECG-specialist zero-shot baselines, LuminaECG improves both waveform measurement and diagnostic recovery. It reaches a clinically meaningful reader tier on the CODE-test benchmark, transfers across geographically diverse ECG datasets without retraining, and generates reports whose structure contains an emergent prognostic signal. These findings suggest that effective ECG agents require not only larger models, but supervision that preserves the alignment between measurable waveform evidence and clinical knowledge.