VLM-in-the-Loop: A Plug-In Quality Assurance Module for ECG Digitization Pipelines

Jiachen Li; Shihao Li; Soovadeep Bakshi; Wei Li; Dongmei Chen

cs.CVApr 1, 2026

VLM-in-the-Loop: A Plug-In Quality Assurance Module for ECG Digitization Pipelines

Authors: Jiachen Li, Shihao Li, Soovadeep Bakshi, Wei Li, Dongmei Chen

Organizations: The University of Texas at Austin

Abstract

ECG digitization could unlock billions of archived clinical records, yet existing methods collapse on real-world images despite strong benchmark numbers. We introduce VLM-in-the-Loop, a plug-in quality assurance module that wraps any digitization backend with closed-loop VLM feedback via a standardized interface, requiring no modification to the underlying digitizer. The core mechanism is tool grounding: anchoring VLM assessment in quantitative evidence from domain-specific signal analysis tools. In a controlled ablation on 200 records with paired ground truth, tool grounding raises verdict consistency from 71% to 89% and doubles fidelity separation ( $Δ$ PCC 0.03 $\rightarrow$ 0.08), with the effect replicating across three VLMs (Claude Opus~~4, GPT-4o, Gemini~~2.5 Pro), confirming a pattern-level rather than model-specific gain. Deployed across four backends, the module improves every one: 29.4% of borderline leads improved on our pipeline; 41.2% of failed limb leads recovered on ECG-Digitiser; valid leads per image doubled on Open-ECG-Digitizer (2.5 $\rightarrow$ 5.8). On 428 real clinical HCM images, the integrated system reaches 98.0% Excellent quality. Both the plug-in architecture and tool-grounding mechanism are domain-parametric, suggesting broader applicability wherever quality criteria are objectively measurable.

Figures & tables

Figure 1: Representative challenging ECGs from the 428-image HCM clinical archive. All three state-of-the-art baselines fail on these images.

Figure 2: System overview. Left: Digitization pipeline with Optimizer Agent closing the loop via 12-action correction space; backend-agnostic interface supports drop-in replacement. Right: Hierarchical QA routing (Tiers 1–3) producing per-lead verdicts. Tool grounding lifts consistency to 89%; 98.0% Excellent on 428 HCM images.

Configuration	Consistency	$\Delta$ PCC
Direct VLM (no rubric)	71% [64%, 77%]	+0.03
VLM + structured rubric	82% [76%, 87%]	+0.05
VLM + rubric + tool evidence	89% [84%, 93%]	+0.08
Full system (+ gate + routing)	91% [86%, 94%]	+0.08

Table 1: Judge ablation on PTB-XL (

n{=}200

). Tool grounding is the key mechanism.

Figure 3: Judge ablation: tool grounding produces the largest gains in consistency and fidelity separation.

Verdict consistency (%)
Configuration	Claude	GPT-4o	Gemini
Direct VLM	71	66	68
+ rubric	82	76	78
+ rubric + tools	89	84	86
Fidelity separation ( $\Delta$ PCC)
Direct VLM	+0.03	+0.02	+0.03

Table 2: Tool-grounding effect across three VLMs (PTB-XL,

n{=}200

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Parameter	Value	Rationale
Rotation	Hough threshold	1200	Detects strong horizontal lines
Rotation	Angle filter	$\pm$ 30 ∘ from horiz.	Excludes non-grid lines
Grid detect.	Red excess	$R{-}(G{+}B)/2>20$	Separates red/pink from gray
Grid detect.	Coverage	$>$ 5% of pixels	Avoids false positives
Grid suppr.	Median kernel	$k{=}21$	Captures local background
Grid suppr.	Desat. divisor	25–80	Adapts to grid saturation

Appendix

Table 3: Pipeline parameters.

Metric	No QA	Tool-only	Full (VLM)
Excellent (%)	93.2	95.8	98.0
Borderline leads	412	338	265
Leads improved	—	74 (6.3%)	147 (12.5%)
Median jitter	0.0015	0.0013	0.0012

Appendix

Table 4: End-to-end ablation on 428 HCM images.

Per-lead
Metric	Med	P90	P95	Max
Jitter ratio	0.0012	0.0029	0.0036	0.0231
Reversal rate	0.0152	0.0256	0.0304	0.0472
Per-image (median across leads)
Jitter ratio	0.0011	0.0026	0.0030	0.0085
Reversal rate	0.0152	0.0248	0.0292	0.0360

Appendix

Table 5: Pipeline quality on 428 HCM images.

Tier routing (5,080 leads)
Component	No VLM	+VLM (all 428)
Tier 1 (SQI screen)	2 (0.0%)
Tier 2 (Morphological)	3,904 (76.9%)
Tier 3 (VLM)	1,174 (23.1%)
Per-image VLM verdicts ( $n{=}428$ )
PASS ( $Q\geq 75$ )	—	162 (37.9%)

Appendix

Table 6: Hierarchical routing on 428 HCM images (

n{=}5{,}080

leads).

Metric	Ours	ECG-Dig.	Open-ECG	ECGMiner
Success rate	428/428	428/428	428/428	0/428
Valid leads	5,049	3,091	1,088	—
Leads/image	11.8	7.2	2.5	—
w/ 12 leads	401	0	0	—
Med. jitter	0.0012	0.0319	0.0194	—
Excellent	98.0%	0.6%	18.7%	—

Appendix

Table 7: Digitization quality on 428 HCM images.

Before VLM feedback
Metric	Ours	ECG-Dig.	Open-ECG	ECGMiner
Leads/image	11.8	7.2	2.5	0.0
Excellent (%)	93.2	0.6	18.7	—
After VLM feedback (up to 3 rounds)
Leads/image	11.9	9.1	5.8	0.0
Excellent (%)	98.0	8.4	31.5	—

Appendix

Table 8: Closed-loop VLM feedback on external backends (428 HCM images).

Generic Action	ECG-Dig. Adapter	Open-ECG Adapter
Expand ROI	Bbox $\times$ 1.3, conf $\downarrow$	Grid margin +20%
Re-threshold	Confidence 0.5 $\to$ 0.3	Otsu offset $-$ 15
Alt. grid detect.	Skip segmentation	Alt. color space
Reprocess lead	Re-run single lead	Re-extract column
Outcome	7.2 $\to$ 9.1 leads	2.5 $\to$ 5.8 leads

Appendix

Table 9: Adapter action mapping (

<

50 lines each).

Backend	Init. deficit	Recovered	Recovery %
Ours	147 / 5,136	147	100%
ECG-Digitiser	2,045 / 5,136	812	39.7%
Open-ECG-Dig.	3,069 / 5,136	1,411	46.0%
ECGMiner	5,136 / 5,136	0	0% (abstain)

Appendix

Table 10: Portability summary: relative lead recovery by backend.

System	VLM	Agent	Loop	Plug	Dims	Ground
MT-Bench ( Zheng et al., 2023 )	–	–	–	–	1	–
MLLM-Judge ( Chen et al., 2024 )	✓	–	–	–	1	–
Prometheus-V ( Lee et al., 2024 )	✓	–	–	–	1	Rubric
AgenticIQA ( Zhu et al., 2025 )	✓	✓	–	–	Multi	–
Self-Refine ( Madaan et al., 2023 )	–	–	✓	–	–	–
VLM-in-the-Loop	✓	✓	✓	✓	4	Tools

Appendix

Table 11: Comparison with related evaluation systems.

Figure 4: VLM quality score distribution (428 HCM images).

VLM Verdict	$n$	PCC (all)	PCC (V1–V6)
PASS ( $Q\geq 75$ )	74	0.33 $\pm$ 0.19	0.48 $\pm$ 0.17
BORDERLINE	124	0.25 $\pm$ 0.21	0.39 $\pm$ 0.19
FAIL ( $Q<40$ )	2	0.11 $\pm$ 0.08	0.18 $\pm$ 0.12
PASS vs. BORDER. $\Delta$ PCC		+0.08 [.03, .13]	+0.09 [.03, .15]

Appendix

Table 12: Fidelity by VLM verdict (PTB-XL,

n{=}200

Lead	PCC	RMSE (mV)	SNR (dB)
I	0.07 $\pm$ 0.14	0.17	$-$ 2.1
II	0.09 $\pm$ 0.12	0.27	$-$ 3.3
III	0.12 $\pm$ 0.14	0.16	$-$ 1.1
aVR	0.20 $\pm$ 0.10	0.17	$-$ 1.9
aVL	0.20 $\pm$ 0.14	0.14	$-$ 2.0
aVF	0.23 $\pm$ 0.14	0.17	$-$ 1.5

Appendix

Table 13: Per-lead fidelity on PTB-XL (

n{=}200

Figure 5: Qualitative examples from PTB-XL (high / median / low fidelity).

Figure 6: Per-lead PCC on PTB-XL (

n{=}200

Figure 7: Fidelity by VLM verdict on PTB-XL ( $n{=}200$ ).

Method	PASS	BORDERLINE	FAIL
SQI-threshold	2.1%	86.6%	11.3%
Jrat-threshold	91.1%	8.7%	0.2%
VLM QA module	37.9%	61.9%	0.2%

Appendix

Table 14: QA verdict comparison (428 HCM images).

Tier	Leads	Latency	Cost/lead
Tier 1 (SQI)	2 (0.04%)	$<$ 1 s	$\sim$ $0
Tier 2 (Morpho.)	3,904 (76.9%)	$\sim$ 5 s	$\sim$ $0
Tier 3 (VLM)	1,174 (23.1%)	$\sim$ 15 s	$\sim$ $0.01

Appendix

Table 15: Routing cost (428 HCM images).

Quality	Expert	VLM
Excellent/PASS	21 (42%)	19 (38%)
Good/BORDERLINE	26 (52%)	30 (60%)
Fair–Poor/FAIL	3 (6%)	1 (2%)
Agreement	72% ( $\kappa=0.48$ [0.29, 0.67])

Appendix

Table 16: Expert vs. VLM assessment (

n{=}50

Context	$w_{\text{SF}}$	$w_{\text{MP}}$	$w_{\text{CU}}$	$w_{\text{VAS}}$	Priority
AF Screening	0.15	0.40	0.35	0.10	P-wave, RR
STEMI	0.15	0.35	0.40	0.10	ST, T-wave
General Archive	0.30	0.25	0.25	0.20	Balanced
Research	0.35	0.30	0.20	0.15	Signal fidelity

Appendix

Table 17: Quality weights by clinical context.

Electrocardiography

Electrocardiography studies the heart’s electrical signals to measure cardiac function and support diagnosis, monitoring, and research.

See all 108 papers

ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment
Jul 6, 2026Xiaocheng Fang, Haoyu Wang, Jieyi Cai +14Time Series Imputation Incomplete Data Imputation
National Institute of Health Data Science, Peking University, Beijing, China · School of Intelligence Science and Technology, Peking University, Beijing, China · Institute of Medical Technology, Peking University Health Science Center, Beijing, China +6
ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation
Mar 15, 2026Jungwoo Oh, Hyunseung Chung, Junhee Lee +6Electrocardiogram Interpretation Clinical Reasoning
KAIST · Ajou University School of Medicine · Yonsei University College of Medicine +1
ELF: A Family of Encoder-Free ECG-Language Models
Jan 5, 2026William Han, Tony Chen, Chaojing Duan +6Healthcare Electrocardiography
Carnegie Mellon University, USA · Allegheny Health Network, USA · University of California, Los Angeles, USA +1

Jul 6, 2026J/K move · Enter open · S save

ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment

Xiaocheng Fang, Haoyu Wang, Jieyi Cai, Qinghao Zhao+13

National Institute of Health Data Science, Peking University, Beijing, China · School of Intelligence Science and Technology, Peking University, Beijing, China · Institute of Medical Technology, Peking University Health Science Center, Beijing, China+6

Time Series Imputation Incomplete Data Imputation Electrocardiography

Open paper AskPDF arXiv

Complete digital 12-lead electrocardiograms (ECGs) are essential for AI-enabled cardiovascular assessment, yet many clinical ECG records, particularly those digitized from ECG images, remain incomplete because of short display formats, incomplete waveform digitization, lead loss, or signal corruption. We developed ImputeECG, a mask-conditioned one-dimensional Transformer autoencoder that completes 12-lead, 10-s ECGs while retaining all observed samples. The model was trained on PTB-XL and evaluated on PTB-XL and CPSC2018 under simulated incomplete settings, with additional real-world validation in a 43,633-record Kailuan clinical cohort after ECG image digitization. Metrics were computed over originally missing regions, with analyses of morphology and downstream diagnostic utility. On PTB-XL, ImputeECG reduced missing-region MAE by 41.7-51.0% and MSE by 54.0-63.7% versus the strongest baseline, with lower errors in R-peak timing, RR interval, QRS duration, QT interval, and P-wave, QRS-complex, and T-wave reconstruction. On CPSC2018, ImputeECG reduced MAE by 49.7-51.9%, supporting external generalization. In downstream multi-label classification, ImputeECG restored performance to 92.28% AUROC and 33.88% AUPRC in the most incomplete PTB-XL setting, approaching complete-ECG performance. On CPSC2018, completed ECGs achieved 94.75-95.89% AUROC and 78.83-81.86% AUPRC across settings. In Kailuan, ECG completion improved zero-shot sex prediction AUROC from 82.6% to 85.8% and reduced age prediction MAE from 10.72 to 9.87 years after image-based ECG digitization. These findings support ECG completion as a practical strategy for converting incomplete ECG records into AI-ready 12-lead, 10-s digital signals and extending the usable scope of ECG archives for digital cardiac assessment.

Abstract

Electrocardiography

Explore similar work

ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment

ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation

ELF: A Family of Encoder-Free ECG-Language Models