PaperlayerView exact revision on arXiv

Source-linked AI summary

ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution

Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Jiangwei Lao, Quanbin Wang, Honglin Lin, Jiajia Liu, Zhihong Lu, Hao Jin, Jian Wang, Lijun Wu

arXiv:2610.17718v1cs.AIcs.LGeess.SPq-bio.QM

TL;DR

ECG models need supervision that targets evolving weaknesses while remaining faithful to waveform evidence. ECG-RSI combines verified waveform synthesis with feedback-guided intervention search; across three rounds, benchmark performance improves, while continued feedback correction yields diminishing external gains.

  • Problem

    It remains difficult to turn a learner’s specific weakness into a controlled ECG example whose waveform, question, and answer agree.

    Source: §1 · p. 1

  • Method

    ECG-RSI verifies requested waveform edits by remeasurement, uses construction errors to repair generation, and uses learner errors to propose training interventions for comparison.

    Source: §1 · pp. 1–2

  • Results

    After three rounds, Qwen3.5-35B-A3B’s mean accuracy across four QA benchmarks rises from 34.57% to 56.04%, and mean F1 across three classification benchmarks rises from 24.27 to 34.03.

    Source: §4.1 · p. 7

  • Takeaways & Limitations

    The retained recipe shifts from direct reinforcement to confusion-focused contrasts, supporting separate diagnosis of weaknesses and selection of interventions within the tested setting.

    Source: §1 · p. 2, §6 · p. 14

    Feedback error falls from 39.63% to 21.33% across R0–R3 while successive external S7 gains shrink to 13.07, 3.01, and 0.38, so feedback-set correction alone does not characterize external progress.

    Source: §5.9 · p. 14

Abstract

from arXiv · show

Electrocardiogram (ECG) models need supervision that targets evolving weaknesses and remains faithful to waveform evidence. We introduce ECG-RSI, a recursive self-improvement framework combining verified waveform synthesis with feedback-guided intervention search. Signal remeasurement verifies waveform edits; construction errors guide generator repair, while learner errors motivate interventions compared on a separate selection set. Across seven benchmarks, three rounds raise Qwen3.5-35B-A3B's mean question-answering accuracy from 34.57% to 56.04% and mean classification F1 from 24.27 to 34.03. Across the retained trajectory, the preferred supervision shifts from direct reinforcement to confusion-focused contrasts. Continued feedback correction yields diminishing external gains. These results motivate separating error diagnosis from intervention selection and checking whether gains on the feedback set transfer to independent benchmarks.

1 Introduction

ECG-RSI addresses the challenge of turning learner weaknesses into controlled ECG examples whose questions and answers remain faithful to waveform evidence. It combines verified waveform synthesis with feedback-guided training interventions.

1 Introduction

  • ECG-RSI separates construction feedback for repairing data generation from learner feedback for proposing and comparing subsequent training interventions.Candidate branches are trained and compared, and the selected branch informs the next round of synthesis and training.Source: §1 · pp. 1–2
  • Target-driven reverse waveform synthesis translates requested ECG feature contrasts into constrained signal edits and verifies the resulting waveforms before supervision is accepted.Remeasurement grounds generated questions and answers in the edited signal.Source: §1 · pp. 1–2
  • The retained trajectory shifts from direct reinforcement to confusion-focused contrasts, while continued feedback correction yields diminishing gains on external benchmarks.The authors distinguish construction feedback, which repairs generation, from learner feedback, which proposes and compares training interventions.Source: §1 · p. 2

2 Related Work

Prior work develops ECG multimodal interpretation, controllable signal synthesis, and feedback-driven self-improvement. ECG-RSI connects these strands by grounding intervention proposals in ECG-specific diagnostics and comparing candidate updates.

2.1 ECG Multimodal Learning and Evidence-Grounded Supervision

  • ECG multimodal work spans instruction tuning, evidence grounding, varied question formats, and datasets, while evaluations probe waveform dependence and stepwise interpretation.ECG-RSI responds by deriving questions and answers from measurements of the resulting signals.Source: §2.1 · pp. 2–3

2.2 Conditional ECG Synthesis and Controlled Interventions

  • Conditional ECG generation ranges from disease- and report-conditioned methods to controllable signals and counterfactuals, while ECG-RSI edits requested feature contrasts and remeasures the waveform.Examples include ECGTwin and TransDiffECG for controllability and counterfactual ECGs for studying prediction changes.Source: §2.2 · p. 3

2.3 Feedback-Driven Data Evolution and Recursive Self-Improvement

  • Recursive self-improvement adapts generated tasks, training data, workflows, or evaluation, and ECG-RSI grounds proposals in ECG-specific diagnostics while separately assessing their utility.Its construction feedback repairs synthesis, learner feedback proposes interventions, and separate selection and external evaluations assess them.Source: §2.3 · p. 3

3 Method

ECG-RSI combines verified ECG question construction with learner-feedback-driven intervention search. Construction feedback repairs synthesis, while learner feedback guides training recipes and candidate selection.

3 Method

  • ECG-RSI uses construction feedback to check waveform and question validity and learner feedback to select data for model-specific utility.Source ECGs and morphology donors are training-only and patient-disjoint from development and independent evaluation data; feedback, selection, and evaluation sets have separate roles.Source: §3 · p. 3, §3.1 · p. 4

3.1 Framework and Feedback Interfaces

  • At each round, construction feedback identifies invalid edits or questions, while learner feedback guides recipes using capability gaps, task type, difficulty, and confusion pairs.Accepted examples and replay data inform training; sibling recipes share a parent, and promotion adds winning examples to the retained pool.Source: §3.1 · p. 4

3.2.1 Targeted Waveform Construction

  • Reverse synthesis edits eligible training ECGs to express a requested feature contrast, then remeasures the signal before deriving supervision.Requests specify target ranges, protected features and tolerances, permitted operators and donors, and an attempt budget; complex diagnoses require morphology beyond interval changes.Source: §3.2.1 · p. 4

3.2.2 Validation and Generator Repair

  • Edited signals must meet target and protected-feature tolerances and pass checks for continuity, lead consistency, morphology, leakage, diversity, and cost before question generation.Generator repairs are matched against prior versions; signal errors revise operators, missing morphology revises retrieval, ambiguity revises questions or context, and unsupported explanations revise prompts.Source: §3.2.2 · pp. 4–5, Figure 2 · p. 5

3.3.1 Learner Diagnosis and Data Recipes

  • Learner errors, invalid outputs, and regressions guide recipes by category, learner-relative difficulty, and supervision type.Training-side probes group examples by reliable success, mixed success, or unresolved status; supervision variants preserve the verified image, question, options, answer, and clinical facts.Source: §3.3.1 · p. 5

3.3.2 Same-Parent Branch Exploration

  • Three same-parent recipe families use direct reinforcement, confusion contrasts, or difficulty-and-task curricula to alter the training-data mixture.Confusion recipes target recurrently confused category pairs and distinguishing ECG evidence; branches add 10,000 unused accepted questions under matched SFT settings.Source: §3.3.2 · pp. 5–6

3.3.3 Model Selection and Retained Lineage

  • After matched training, the highest-scoring candidate that meets output-validity and non-target regression tolerances is retained; otherwise, the parent remains.The retained learner is reevaluated on the feedback set and recipe directions are chosen again each round.Source: §3.3.3 · p. 6

4 Experiments

The experiments evaluate ECG-RSI across ECG tasks and track benchmark, validation, and feedback results over successive rounds. ECG-RSI R3 leads the seven-benchmark aggregate comparison, while round-wise results show overall gains alongside task-specific fluctuations.

4 Experiments

  • The evaluation asks how feedback granularity and candidate directions affect training utility and transfer across eight ECG tasks totaling 10,840 items.Source: §4 · p. 6

4.1 Main Benchmark Results

  • After three rounds, Qwen3.5-35B-A3B mean QA accuracy increases from 34.57% to 56.04%, and mean classification F1 from 24.27 to 34.03.Source: §4.1 · p. 7
  • 46.61 S7 gives ECG-RSI R3 the highest aggregate score among compared models across seven benchmarks.Source: Table 1 · p. 7, §4.1 · p. 7

4.2 Evolution across Rounds

  • Each winning R-A branch adds 10,000 accepted questions and replays prior data; aggregate performance rises through R3, though several metrics peak at R2.Source: Table 2 · p. 7, §4.2 · p. 7

4.3 Validation-Set Performance

  • MedX accuracy rises from 51.57% to 57.24% despite a small R2 decline, while LUDB rises every round from 53.45% to 67.24%.Source: Table 3 · p. 8, §4.3 · p. 8

4.4 Feedback Diagnostics across Rounds

5 Ablation Studies

The ablations assess waveform and supervision quality, then compare how feedback granularity, intervention choice, and evolution strategy affect training utility and external performance. Across rounds, the preferred supervision changes, while continued feedback-error reduction brings diminishing external gains.

5.1 Effectiveness of the Synthesis Method

  • All six synthesized waveform features meet the adopted MAE criteria in a single run of 10,000 samples.The reported validation includes MAE and RMSE; the criteria are experiment-specific.Source: §5.1 · pp. 9–10, Table 6 · p. 10

5.2 Blinded Review by Human Physicians

  • Mean physician ratings are 0.88 for answer correctness and 0.76 for explanation support, with κ = 0.78.Target attainment ranges from 0.78 for RR to 0.85 for HR; explanation support ranges from 0.72 to 0.78 and remains below answer accuracy for every edit type.Source: §5.2 · pp. 10–11, Table 7 · p. 10, Table 8 · p. 10

5.3 Multi-Model Question-Quality Scores

  • The full-method Question score is 85.09, the unweighted mean across five models and 23 ECG categories.R1 uses 10,000 of 16,193 candidate questions.Source: Figure 5 · p. 11, §5.3 · p. 11

5.4 Data Construction and Feedback Selection; 5.5 Synthesis-Component Ablation

  • Adding reverse synthesis raises S7 to 45.56, and feedback-guided selection raises it to 46.23; freezing synthesis updates increases candidates needed and reduces quality and downstream performance.At the fixed target of 10,000 accepted questions, the full method requires 16,193 candidates, versus 36,723 with operator updates frozen and 22,398 with skill updates frozen.Source: §5.4 · p. 11, §5.5 · p. 11, Table 9 · p. 11

5.7 Round-Wise Branch Selection and Data Composition

  • Branch A beats the contrast-focused branch in the first comparison, with S_sel = 42.66 versus 40.51, despite a 52% Mobitz I/II confusion error.Later comparisons favor regional STEMI contrasts at 45.08 and rhythm contrasts at 46.03; curriculum is not selected in either comparison despite persistent L3 errors.Source: §5.7 · pp. 12–13, Table 12 · p. 12, Table 13 · p. 13

5.8 Comparison of RSI Evolution Strategies

5.6 Feedback Granularity and Training Utility

  • With 10,000 added questions, full feedback reaches S7 = 46.23, versus 45.56 with fixed selection and 45.60 with category-only feedback.Adding question type raises S7 to 45.98, and including difficulty and confusion pairs raises it to 46.23.Source: §5.6 · p. 12, Table 11 · p. 12, §5.9 · p. 13

5.9 Results Analysis

  • Feedback error falls from 39.63% to 21.33% across R0–R3, while successive external S7 gains shrink to 13.07, 3.01, and 0.38.ECG-QA, CPSC2018, and CODE-15% regress in R3 despite fewer feedback errors; the retained path shifts from direct reinforcement to confusion-focused supervision.Source: §5.9 · pp. 13–14

6 Conclusion

ECG-RSI combines verified waveform synthesis with feedback-guided intervention search, shifting the retained recipe from direct reinforcement to confusion-focused contrasts. Across the tested setting, continued feedback correction yields diminishing external gains, supporting separate weakness diagnosis and intervention selection.

6 Conclusion

  • Across three rounds, S7 rises from 30.15 to 46.61, and MedXpert ECG accuracy from 51.57% to 57.24%.The retained recipe shifts from direct reinforcement to confusion-focused contrasts; continued feedback-error reductions yield diminishing S7 gains, so transfer and regression checks remain necessary.Source: §6 · p. 14

Appendix Contents and Reading Guide

  • The appendix guide covers supplementary feedback counts, branch evaluations, task taxonomy, waveform construction, prompts, and training configuration.It also lists multi-seed results, evaluation splits, and parameter extraction and selection details.Source: Appendix Contents and Reading Guide · p. 18

A.1 Student-Conditioned Data Recipes

  • Recipes combine task category, learner-relative difficulty, and supervision representation, with constraints, replay, budgets, and a versioned operator and prompt library.Difficulty is estimated from repeated responses by the current learner; supervision may use teacher, evidence-expanded, or error-targeted rationales.Source: §A.1 · pp. 18–19

A.2 Construction Feedback and Versioned Repair

  • Construction failures route repairs to the relevant operator, applicability or retrieval rule, question/context, or prompt, followed by regression checks.Repairs are compared on the same source signals, requests, random seeds, and attempt budgets before adoption.Source: §A.2 · p. 19

A.3 Waveform Construction and Validation Details

  • Bidirectional synthesis either derives questions from measured real waveforms or edits eligible training waveforms to a requested contrast and remeasures them.Edits must achieve the target while preserving specified non-target properties; sources and donors are patient-disjoint from feedback, selection, and evaluation.Source: §A.3 · p. 19

A.4 Bounded Recipe Search within the Learning Loop

  • Each round probes learner errors, proposes bounded recipes, validates data, compares matched branches from the same parent, and promotes a qualifying candidate.Selection uses a separate set from external S7, and the retained path continues from the promoted model while unselected branches are archived.Source: §A.4 · p. 20

A.5 Training Configuration

Training uses a fixed configuration with the language model trainable and the vision encoder and multimodal aligner frozen. The parameter-extraction appendix describes ECG measurement scales and selection based on reliability and diagnostic utility.

A.5 Training Configuration

  • Training uses three epochs, a maximum sequence length of 28,672 tokens, and a trainable language model with the vision encoder and multimodal aligner frozen.Distributed Adam uses cosine learning-rate decay with 5% warmup; computation uses bfloat16 while master parameters, gradients, and Adam states use float32.Source: §A.5 · p. 20

A.6 Parameter Extraction and Selection Details

  • At 25 mm/s, 1 mm horizontally represents 40 ms; at 10 mm/mV, 1 mm vertically represents 0.1 mV, and parameter selection uses reliability and diagnostic utility.The appendix also describes a common dimensionless normalization for comparing inverse error scales across bpm, ms, and mV measurements.Source: §A.6 · p. 21

B Supplementary Experiments and Data

Supplementary experiments report round-wise learner feedback and difficulty-level error trajectories, alongside category-level records across 23 ECG task categories.

B Supplementary Experiments and Data

B.1 Difficulty-Level Feedback Counts

  • L3 error falls from 58.20% to 38.17%, the largest reduction across difficulty levels, but remains the highest final error rate.L1 falls to 12.97%; the table reports error counts and rates with group sizes in the header.Source: §B.1 · p. 21, Table 17 · p. 21

B.2 Category-Level Feedback Counts

  • Category-level records report error counts and rates for 23 ECG categories across R0–R3, with each question assigned to one primary category.The category counts sum to 3,000, and each rate divides its error count by the corresponding group size.Source: Table 18 · p. 21

B.3 Supplementary Branch-Evaluation Records

The branch records compare same-parent candidates on a shared R0→R1 development protocol and identify the promoted candidate. The reported three-round retained path is A→B→B.

B.3 Supplementary Branch-Evaluation Records

  • The R0→R1 comparison reports overall accuracy, three macro scores, and selection score S, with ΔS measured relative to the parent.Candidates use the same development protocol and each adds 10,000 accepted questions with matched historical replay and training settings.Source: §B.3 · p. 22
  • Tab. 12 reports all three rounds of candidate selection and the full retained path A→B→B.Source: §B.3 · p. 22
  • All candidates pass the score and decline gates, and A is promoted with the highest selection score (42.66).B and C remain eligible; worst-group changes are accuracy differences from the parent in percentage points.Source: Table 19 · p. 22, §B.3 · p. 22

B.4 Multi-Seed Results on ECG Tasks

Across three training runs, mean ECG-task scores rise from R1 to R3, with low reported variability across trained rounds.

B.4 Multi-Seed Results on ECG Tasks

  • Scores are reported for three independent training runs, with SD defined as the sample standard deviation.The base row uses the same fixed initialization.Source: §B.4 · p. 23
  • Mean ECG-task scores rise from 33.40 at R1 to 37.82 at R3.The reported SD across trained rounds is 0.22–0.29.Source: §B.4 · p. 23

B.5 Evaluation Splits

The evaluation reports enabled splits, distinguishes the first seven from MedXpert-ECG, and maps abbreviated figure labels to dataset names.

B.5 Evaluation Splits

  • The first seven enabled evaluation splits contribute equally to S7, while MedXpert-ECG is reported separately.Tab. 21 reports evaluation-item counts.Source: Table 21 · p. 23
  • The listed configurations exclude ptb-test-report (500 items) and arena (50 items), and LUDB is evaluated separately from the eight enabled splits.MMMU-ECG, G12EC, ECG-QA, PTB-XL, CPSC2018, CODE-15%, and MedXpert-ECG are described in the passage.Source: §B.5 · p. 23
  • Figure labels MMMU, G12, ECGQA, PTB, CPSC, and CODE15 denote MMMU-ECG, G12EC, ECG-QA, PTB-XL, CPSC2018, and CODE-15%, respectively.Source: §B.5 · p. 23

C ECG Task Taxonomy

Appendix C organizes ECG task categories; Appendix D documents reusable stage-wise prompt templates with request-specific fields and tool results.

C ECG Task Taxonomy

  • The taxonomy groups 23 ECG task categories into related medical categories, with sector areas and counts representing category numbers.It summarizes task coverage and the grouping used for category-level feedback.Source: C ECG Task Taxonomy · p. 23, Figure 6 · p. 24

D.1 Constrained waveform construction

  • Waveform construction checks source and landmark reliability, applies constrained edits, remeasures targets and protected-feature drift, and repairs or rejects failed edits.The output records measurements, validity checks, and failure codes.Source: §D.1 · p. 24

D.2 Evidence-grounded question and context generation

  • Question generation uses verified ECG evidence to create one answerable question with the requested options, plausible incorrect distractors, and explicitly marked synthetic teaching context when used.Questions stay within the supported measurement, pattern, or diagnosis level.Source: §D.2 · p. 25

D.3 Teacher answer and evidence explanation

  • The teacher-answer template requests step-by-step reasoning and a final answer in boxed form.Source: §D.3 · p. 25

D.4 Independent quality and acceptance review

  • Independent review solves the question and audits waveform, answer, context, leakage, and explanation gates; acceptance requires both an aggregate score of at least 80/100 and every mandatory gate passing.The reviewer returns gate-wise decisions and supporting evidence.Source: §D.4 · p. 25

D.5 Construction failure and versioned repair

  • Construction repairs map observed failure types to specific component changes and compare old and new versions on matched sources, requests, seeds, and budgets.Exhausted repair budgets lead to quarantine or rejection.Source: §D.5 · p. 25

D.6 Learner-feedback diagnosis and recipe proposal

  • Learner-feedback diagnosis proposes distinct recipes for direct weakness coverage, confusion contrasts, and difficulty or evidence-granularity progression.Candidates retain the same cumulative training set and each add 10,000 first-use accepted questions; only the winner's additions enter the next cumulative set.Source: §D.6 · p. 26

D.7 Candidate comparison and model promotion

  • Promotion compares eligible candidates under the supplied development-only score, regression, and valid-output gates, otherwise retaining the parent and recording admitted additions and replay provenance.Comparability checks include parent, training budget, and evaluation protocol.Source: §D.7 · p. 26

E Prompts and Case Studies

The appendix presents two constrained ECG edits and corresponding parameter questions, with saved responses illustrating the intended measurements and interpretations.

E Prompts and Case Studies

Case 1: PR prolongation

  • For PR prolongation, the edit lengthens PQ and shortens TP to preserve RR; the target PR is 300 ms and remeasurement gives 298 ms.The accompanying figure reports the same target and remeasured PR.Source: Case 1: PR prolongation · p. 26, Figure 7 · p. 27
  • The retained answer identifies HR 74.6 bpm, PR 298 ms, QRS 96 ms, and QTc 393.7 ms as first-degree AV block.This matches option C in the displayed choices.Source: Case 1: PR prolongation · pp. 26, 28

Case 2: Heart-rate slowing

  • For heart-rate slowing, the edit extends TP while copying P–QRS–T; the target rate is 48 bpm and remeasurement gives 47.9 bpm.The figure describes the same target and remeasured rate.Source: Figure 8 · p. 29, Case 2: Heart-rate slowing · p. 30
  • The retained answer identifies HR 47.9 bpm, PR 150 ms, QRS 48 ms, and QTc 348.5 ms as sinus bradycardia with otherwise normal intervals.The response selects option B.Source: Case 2: Heart-rate slowing · p. 31
  • The Case 2 error analysis notes that eight beats in ten seconds give approximately 48 bpm, using seven R–R intervals from first to last peak.It also notes that the response invokes the synthetic origin when justifying QRS duration.Source: Case 2: Heart-rate slowing · p. 31

Related papers

Loading 2610.17718v1…