ECG Diagnostic Evidence Rewards

• ECG Diagnostic Evidence Rewards are a set of training objectives that assign value to model outputs based on alignment with measurable diagnostic ECG features.
• They employ methods like reinforcement learning, semantic recovery scores, and kappa-based metrics to preserve clinical evidence within predictions.
• These rewards improve model robustness and interpretability by ensuring that ECG predictions remain tightly anchored to clinically relevant waveform data.

Searching arXiv for the cited ECG diagnostic evidence reward papers to ground the article. {"query":"(Jin et al., 4 Feb 2026) ECG-R1 Protocol-Guided and Modality-Agnostic MLLM for Reliable ECG Interpretation", "max_results": 5} {"query":"(Alcaraz et al., 7 Feb 2025) Explainable and externally validated machine learning for neuropsychiatric diagnosis via electrocardiograms", "max_results": 5} “ECG Diagnostic Evidence Rewards” denotes a class of training objectives, evaluation signals, and evidence-grounding mechanisms that assign value to an ECG model in proportion to how strongly its predictions or explanations are supported by diagnostically meaningful waveform information. In current work, the concept appears in several forms: explicit reinforcement-learning rewards over stepwise ECG evidence and diagnosis accuracy, semantic recovery scores under adversarial missingness, kappa-based preservation of diagnostic morphology after ECG processing, and cross-modal or causal objectives that force ECG representations or language outputs to remain anchored to clinically relevant evidence rather than merely plausible text or high aggregate accuracy (Jin et al., 4 Feb 2026).

1. Conceptual scope and taxonomy

The literature separates explicit reward design from implicit evidence shaping. The clearest explicit formulation appears in ECG-R1, which uses “Reinforcement Learning with ECG Diagnostic Evidence Rewards” to reward a multimodal LLM for producing protocol-structured, evidence-grounded ECG interpretations. Its total reward sums a format term, a final-diagnosis Jaccard term, and a process reward that measures stepwise coverage of key evidence phrases extracted from a protocol-guided reference (Jin et al., 4 Feb 2026).

Other systems implement functionally similar mechanisms without always naming them as rewards. SCAR treats robustness as preservation of ECG–text semantics when alignment-critical spatio-temporal evidence is removed, and operationalizes “diagnostic evidence rewards” through semantic recovery and the Counterfactual Missingness Resolution Score (CMRS) (Liu et al., 2 Apr 2026). CARE-ECG exposes latent biomarkers, causal evidence paths, hallucination risk, and retrieval-grounded explanation signals that the paper states can be converted into explicit rewards for RLHF, RLAIF, or self-verification training (Khatibi et al., 12 Apr 2026). By contrast, ALFRED does not define an explicit reward function or reinforcement-learning mechanism; instead, it relies on surrogate evidence-quality signals such as cosine-similarity retrieval relevance, rule satisfaction, and expert curation (Yu et al., 30 Apr 2025).

A further branch of the topic concerns evidence preservation rather than evidence generation. “Diagnostic Quality Assessment for Low-Dimensional ECG Representations” defines diagnostic resemblance as beyond-chance agreement between labels assigned to original and processed ECGs, using kappa as the primary measure of preserved diagnostic evidence (Kovács et al., 2022). Multimodal ECG–CMR pretraining similarly uses masked data modeling and multimodal contrastive alignment as evidence-alignment signals, rewarding ECG embeddings that encode structure and function normally observed in cardiac magnetic resonance imaging (Turgut et al., 2023).

Approach	Evidence substrate	Reward or scoring role
ECG-R1	Protocol-guided key phrases, diagnosis sets, format	Explicit RL reward
SCAR	Alignment-critical tokens, secondary cues, CMRS	Robustness reward under missingness
CARE-ECG	Causal drivers, retrieved facts, hallucination rate	Reward-capable causal signals
ALFRED	Retrieval cosine similarity, rule flags, expert knowledge	Surrogate evidence-quality signals
Diagnostic quality assessment	Original vs processed ECG labels	Kappa-based preservation score
MMCL + MDM	Masked ECG content and paired CMR	Cross-modal evidence alignment

This taxonomy suggests that the topic is broader than reinforcement learning narrowly construed. In the surveyed literature, “reward” may denote an explicit scalar return, a contrastive or reconstruction objective, a semantic preservation score, or a chance-corrected measure of diagnostic fidelity.

2. What counts as diagnostic evidence

Across the cited work, diagnostic evidence is always tied to measurable or inferable ECG structure rather than free-form textual plausibility. In ECG-R1, FeatureDB deterministically measures per-beat and per-lead evidence from the time series, yielding 14 sequences per lead over 12 leads, including heart rate, RR intervals, P amplitude and duration, PR interval, QRS amplitude and duration, T amplitude and duration, ST descriptors, and QT/QTc intervals. These features are paired with protocol-derived thresholds and logic such as bradycardia $< 60$ bpm, normal PR $0.12$–$0.20$ s, prolonged PR $> 0.20$ s, normal QRS $< 0.10$ s, bundle branch block $> 0.12$ s, left-axis deviation if Lead I is positive and Lead II negative, and ST-elevation thresholds of at least $1$ mm in limb leads or $2$ mm in precordial leads (Jin et al., 4 Feb 2026).

ALFRED uses a different but comparably structured evidence substrate. A UNet-based delineation model segments raw 12-lead ECGs and yields 30 lead-specific features per lead, 12 global features, and subject metadata. An expert rule module converts those measurements into True/False diagnostic flags for 40 categories, including ECG quality warnings and common conditions such as bundle branch block and atrial fibrillation. Retrieval is then driven by feature names and diagnosis names so that the generated explanation remains linked to the measured features and the rule outputs (Yu et al., 30 Apr 2025).

In SCAR, evidence is localized at the token level. A multi-lead ECG $X \in \mathbb{R}^{L \times C}$ is partitioned into spatio-temporal tokens, and “primary diagnostic evidence” is defined as the lead–time morphology most responsible for ECG–text alignment. The adversarial masker identifies alignment-critical tokens whose removal maximally disrupts cross-modal similarity, while the selector learns to upweight “secondary yet diagnostically informative morphological cues” that remain visible after corruption (Liu et al., 2 Apr 2026).

CARE-ECG reframes evidence in causal terms. A 12-lead, 10 s ECG segment $x \in \mathbb{R}^{C \times T}$ with $0.12$0 and 500 Hz sampling is encoded into temporally organized latent biomarkers $0.12$1, discretized into ordinal regimes $0.12$2, and connected to diagnoses $0.12$3 through a Bayesian causal graph. Retrieved medical facts and the final explanation are conditioned on these latent drivers and posterior hypotheses, so evidence becomes a chain from waveform to biomarker to causal posterior to cited fact (Khatibi et al., 12 Apr 2026).

In externally validated neuropsychiatric diagnosis from ECG, evidence is a harmonized feature vector comprising RR-interval, PR-interval, QRS-duration, QT-interval, QTc-interval, P-wave axis, QRS axis, T-wave axis, age, and sex. The model’s evidential interpretation is provided by Shapley values rather than by a rule engine or language reward (Alcaraz et al., 7 Feb 2025). In the P-wave/PR-interval Q-learning study, diagnostically salient evidence is concentrated in Leads II and V1 and includes P-wave amplitude, duration, polarity, morphology flags, PR interval, and class-specific indicators such as PR $0.12$4 ms or P-wave absence (Fatima et al., 2024).

3. Explicit reward formulations

ECG-R1 provides the most concrete implemented definition. For a protocol-structured target response $0.12$5 with $0.12$6 steps and a generated response $0.12$7, the step reward is the fraction of extracted key evidence phrases recovered in the corresponding generated step:

$0.12$8

and the process reward is

$0.12$9

Final-diagnosis accuracy is the Jaccard similarity between predicted and gold diagnosis sets, and the total reward is

$0.20$0

The policy is optimized with DAPO, a decoupled-clipping PPO variant with $0.20$1, $0.20$2, and $0.20$3 (Jin et al., 4 Feb 2026).

SCAR defines evidence reward in semantic terms. Its training objective combines adversarial removal, full-view consistency, and compensation:

$0.20$4

The paper then makes the “evidence reward” explicit as incremental semantic recovery after compensation,

$0.20$5

or via normalized alignment or CMRS. This reward is not merely diagnostic correctness; it measures whether the model recovers the full-view diagnostic semantics after the primary evidence has been removed (Liu et al., 2 Apr 2026).

The Q-learning study on P waves and PR intervals implements reinforcement learning directly at the beat-classification level. The update rule is

$0.20$6

with $0.20$7, $0.20$8, and SoftMax temperature $0.20$9. The paper evaluates three reward variants—correct classification, time-penalized reward, and confidence-augmented reward—and then describes an evidence-driven shaping term that penalizes PR deviation, P-wave detection error, and poor signal quality while rewarding confidence and correctness (Fatima et al., 2024).

The diagnostic-preservation framework uses kappa as the core reward-like signal:

$> 0.20$0

Here, $> 0.20$1 is observed agreement and $> 0.20$2 is chance agreement derived from the marginals. This converts “diagnostic evidence” into a chance-corrected preservation score: a processing method is rewarded when the original and processed ECGs retain the same categorical diagnostic morphology beyond what could occur by chance alone (Kovács et al., 2022).

CARE-ECG presents explicit reward design as an extension rather than as the paper’s implemented training loss. Its proposed composite reward includes diagnostic correctness, faithfulness, causal consistency, and hallucination penalties:

$> 0.20$3

This is notable because it treats evidence reward as a property of causal explanation quality rather than only label accuracy (Khatibi et al., 12 Apr 2026).

4. Evidence grounding beyond explicit rewards

Not all evidence-grounded ECG systems optimize an explicit reward, yet several clearly impose evidence-aligned constraints. ALFRED is the clearest counterexample to the assumption that reward design is necessary for evidence grounding. The paper states that it does not use reinforcement learning, a formal reward function, or a combined scoring equation. Reliability instead emerges from structured prompt composition: quantitative ECG features, 40 rule results marked True/False, retrieved feature definitions, and retrieved diagnosis definitions. The paper reports that adding rule results produces the largest jump across PPV, NPV, sensitivity, and specificity, indicating that rule outputs function as strong diagnostic anchors even without an explicit reward (Yu et al., 30 Apr 2025).

The neuropsychiatric ECG study likewise does not implement reward learning, but it provides a different form of diagnostic evidence. One binary XGBoost classifier is trained per ICD-10 condition, AUROC is reported on both MIMIC-IV-ECG and ECG-ViEW II, and Shapley values quantify the contribution of each ECG feature. Age is the most important predictor across conditions except anoxic brain damage; QTc-interval and T-wave axis dominate Parkinson’s disease and anoxic brain damage; RR-interval and PR-interval dominate unspecified dementia and delirium. The paper argues that these feature attributions align with known markers while also suggesting potentially new markers such as high RR-interval for Alzheimer’s disease, low QRS axis for Parkinson’s disease, and low QRS duration for vascular dementia (Alcaraz et al., 7 Feb 2025).

The ECG–CMR transfer paper places evidence alignment at the representation level. Masked data modeling on ECG imposes a reconstruction penalty under a high masking ratio $> 0.20$4, while multimodal contrastive learning aligns ECG embeddings with paired CMR embeddings through a CLIP-style loss with $> 0.20$5 and $> 0.20$6. The paper interprets this as a cross-modal “evidence reward”: ECG representations are rewarded when they encode CMR-relevant cardiac structure and function strongly enough to identify the correct paired CMR among batch negatives (Turgut et al., 2023).

These systems clarify a common misconception: evidence grounding is not reducible to post hoc explainability. In the surveyed literature, evidence may be enforced by retrieval structure, contrastive pairing, consistency constraints, feature attribution, or diagnostic-preservation statistics before it is ever expressed as a natural-language explanation.

5. Empirical behavior and reported gains

Implemented evidence rewards can improve both correctness and grounding. On the grounded ECG interpretation benchmark of 2,381 ECGs from ECG-Grounding, ECG-R1’s reinforcement-learning stage improves Diagnosis Accuracy from 79.33 to 80.29, Analysis Completeness from 6.36 to 6.51, Analysis Relevance from 4.58 to 4.74, Lead Evidence Validity from 5.53 to 5.81, ECG Feature Grounding from 79.92 to 80.57, Evidence-Based Reasoning from 78.08 to 79.08, and Clinical Diagnostic Fidelity from 83.51 to 84.20. An ablation removing EDER reduces process-quality dimensions more substantially than final diagnosis, indicating that the reward especially shapes intermediate reasoning quality (Jin et al., 4 Feb 2026).

SCAR reports robustness gains under joint lead and temporal missingness across six downstream tasks. Under hard missingness, CMRS reaches 79.26 on CSN versus at most 42.77 for baselines, 68.45 on PTBXL-Rhythm versus at most 48.66, and 74.40 on PTBXL-Super versus at most 38.12. The paper explicitly argues that CMRS reflects semantic preservation better than AUROC alone. It also reports stronger linear-probing transferability, with PTBXL-Rhythm AUROC of 93.09, 96.28, and 97.02 at 1%, 10%, and 100% labels (Liu et al., 2 Apr 2026).

In externally validated neuropsychiatric diagnosis, the evidentially interpretable feature model attains strong discrimination despite markedly lower prevalence in the external cohort. Alzheimer’s disease $> 0.20$7 reaches an internal AUROC of 0.8134 $> 0.20$8 and an external AUROC of 0.8678 $> 0.20$9; unspecified dementia $< 0.10$0 reaches 0.8490 $< 0.10$1 internally and 0.8620 $< 0.10$2 externally. Parkinson’s disease, anoxic brain damage, vascular dementia, and delirium all remain above AUROC 0.7 in the reported results (Alcaraz et al., 7 Feb 2025).

ALFRED demonstrates that evidence-quality signals can improve zero-shot ECG diagnosis even without explicit RL. On PTB-XL diagnostic superclasses, the Base configuration yields PPV 0.326, NPV 0.805, Sensitivity 0.356, and Specificity 0.754; the Proposed configuration yields PPV 0.443, NPV 0.834, Sensitivity 0.477, and Specificity 0.797. The largest improvement comes from adding rule results, while adding expert knowledge further improves performance, especially in Ablation3 for PPV and Specificity (Yu et al., 30 Apr 2025).

The P-wave/PR-interval Q-learning study reports average accuracy of 90.4%, average Hamming loss of 9.6%, average classification time of 0.04 s at the 100th episode with around 40,000 samples, and average training reward of 344.05 at $< 0.10$3, $< 0.10$4, and $< 0.10$5. The ECG–CMR transfer study reports that MMCL+MDM improves CVD risk prediction by up to 12.19% and phenotype prediction by up to 27.59%, with CAD, AF, and DM AUROCs of $< 0.10$6, $< 0.10$7, and $< 0.10$8 respectively for ECG-only inference after pretraining (Fatima et al., 2024, Turgut et al., 2023).

The aggregate pattern is consistent: evidence-aware objectives tend to show their clearest advantages when standard performance metrics are insufficient by themselves—under missingness, cross-modality transfer, explanation faithfulness, external validation, or clinically structured reasoning.

6. Limitations, controversies, and open directions

A central controversy is how evidence should be measured. ECG-R1 uses exact phrase matching against up to three short phrases per step, which makes the reward operationally simple but can fail on clinically valid paraphrases. The paper also notes that unsupported claims are only indirectly penalized: they fail to earn phrase matches and may reduce Jaccard diagnosis overlap, but there is no explicit negative reward for hallucinated lead-level assertions (Jin et al., 4 Feb 2026).

SCAR’s CMRS depends on the choice of reference classifier $< 0.10$9, and the paper notes failure cases when both primary and most secondary cues are jointly removed or heavily corrupted. CARE-ECG raises a related issue at the causal layer: graph structure learned from observational data may contain spurious edges or discretization artifacts, so reward design based on causal paths can inherit errors from the graph itself. The paper therefore proposes interventional rewards and counterfactual consistency checks as mitigation rather than presenting the problem as solved (Liu et al., 2 Apr 2026, Khatibi et al., 12 Apr 2026).

In rule-grounded RAG, another limitation is that evidence availability does not imply fully specified diagnostic logic. ALFRED states that exact threshold values and decision rules are not enumerated in the paper, calibration metrics such as AUROC and ECE are not reported, and the full proposed prompt can slightly reduce PPV and Specificity relative to Ablation3, likely because increased prompt size confuses the LLM (Yu et al., 30 Apr 2025).

Evidence preservation frameworks face different statistical complications. The kappa-based diagnostic-quality paper emphasizes the prevalence paradox: high observed agreement can still yield low $> 0.12$0 when prevalence is extreme because chance agreement is also high. It also documents cases in which objective distortion measures such as PRD, WWPRD, and WEDD contradict human assessment of diagnostic preservation, particularly when noise overlaps diagnostically relevant frequency bands (Kovács et al., 2022).

Finally, externally validated and explainable ECG diagnosis still faces unresolved confounding. In neuropsychiatric prediction, therapy-related cardiotoxicity from acetylcholinesterase inhibitors, dopamine agonists, and antipsychotics can blur the distinction between disease-related and treatment-related ECG changes, and acquisition or preprocessing details, calibration, missing-data handling, and subgroup fairness analyses are not reported (Alcaraz et al., 7 Feb 2025). This suggests that future “ECG Diagnostic Evidence Rewards” will likely need to couple evidential grounding with confounder control, calibration, and broader external validation rather than treating evidence coverage alone as sufficient.