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

doi:10.48550/arXiv.2610.17718

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

Xiaoliang Wang^1,2,* Zhuo Chen^2,* Chuanyang Zheng² Jiangwei Lao² Quanbin Wang² Honglin Lin^3,4 Jiajia Liu² Zhihong Lu² Hao Jin¹ Jian Wang² Lijun Wu^2,3,†

¹ Zhejiang University
² Ant Group
³ Shanghai Artificial Intelligence Laboratory
⁴ Shanghai Jiao Tong University

* Equal contribution† Corresponding author

Advancing ECG reasoning through waveform-grounded self-improvement.

Paper ↗GitHub ↗

ECG evidence → targeted supervision

Independent ECG evaluation

Schematic08 / 08

Signal construction / verification

· PR target· R–R + QRS· Answer groundingWaveform checks → construction repair

ECG questionWhich interval changed?

PR contrasts at matched R–R; verification is pending.

Awaiting checks

Candidate ECG Q/AECG learner

Development loop

ECG synthesis agent

Retained

Waveform: ~~Stretch PQ~~PQ + TP compensation
Teaching: ~~Direct practice~~PR contrast pairs
Focus: ~~Weak ECG categories~~Interval confusion

Development feedback

Interval confusionContrast recipes

Learner assessment pending

Retained branch

Retained learner → external evaluation

Development feedback

Retained learnerIndependent evaluation

7 benchmarks · MedXpert-ECG · LUDB

Held out from generation and selection

Schematic ECGs · feedback and selection use separate development sets.

Teaching choices and transfer

The preferred intervention changes.

R1Reinforce

R2Contrast

R3Contrast

Development feedback proposes interventions; separate selection compares their utility.

External gains diminish.

R1+13.07

R2+3.01

R3+0.38

Seven-benchmark mean gain (points).

+21.47ppMean QA accuracy

+9.77pointsMean classification F1

Qwen3.5-35B-A3B · three rounds
Seven benchmarks · reported R-A trajectory

Original paper Figure 4: question-answering accuracy, classification F1, validation accuracy, and tree, uniform-allocation and chain strategy trajectories across R0 to R3. — Figure 4. Accuracy, F1, validation and strategy comparisons. S7 is the equal-weight mean of seven benchmark scores. Scroll to compare panels.

Experiment details and benchmark scores

S7 equally weights four accuracy and three classification F1 scores on a 0–100 scale. The main trajectory rises from 30.15 to 46.61. Mean QA accuracy increases from 34.57% to 56.04%; mean F1 increases from 24.27 to 34.03. Gains are computed before rounding.

ECG-RSI main trajectory (R-A)
Benchmark	Metric	R0	R1	R2	R3
MMMU-ECG	Accuracy	42.00	48.00	56.00	57.50
CSN	Accuracy	30.91	61.68	64.86	67.23
G12EC	Accuracy	41.36	64.61	66.39	66.24
ECG-QA	Exact-match accuracy	23.99	33.56	34.55	33.18
PTB-XL	F1	29.50	32.90	36.40	38.30
CPSC2018	F1	13.10	23.00	23.90	22.80
CODE-15%	F1	30.20	38.80	41.50	41.00
S7	Equal-weight mean	30.15	43.22	46.23	46.61

Interventions are compared from the same parent on a separate development selection set. External benchmarks remain held out from generation and selection. Source: the paper’s main results and “Round-Wise Branch Selection and Data Composition.”

Recorded waveform example and full method

The appendix PR example requests 300 ms and remeasures 298 ms, with heart rate 74.6 bpm. It lengthens the PQ segment and compensates in the TP segment to preserve the RR interval. These values follow the paper’s automatic measurement convention.

Recorded PR-prolongation example from the paper — Paper appendix, Case 1. Click to inspect the original and edited ECGs.

ECG-RSI framework with construction and learner feedback loops — Paper Fig. 1. Construction feedback repairs operators and prompts. Development feedback proposes recipes; a separate development set selects a model from same-parent candidates.

Paper ↗Project ↗GitHub ↗Related: MedEvo ↗

Research code, model weights and training data are not released.

Citation

@misc{wang2026ecgrsi,
  title = {{ECG-RSI}: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution},
  author = {Wang, Xiaoliang and Chen, Zhuo and Zheng, Chuanyang and Lao, Jiangwei and Wang, Quanbin and Lin, Honglin and Liu, Jiajia and Lu, Zhihong and Jin, Hao and Wang, Jian and Wu, Lijun},
  year = {2026},
  eprint = {2610.17718},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi = {10.48550/arXiv.2610.17718},
  url = {https://arxiv.org/abs/2610.17718}
}