Repeat variation and definition dependence of ECG/VCG markers: an exploratory audit of one implementation (LoopLens 10.1)

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Routine diagnostic ECG/VCG evaluation (n=10187). Intervention: LoopLens 10.1 ECG/VCG marker pipeline (repeat variation, Kors vs inverse Dower reconstruction matrix, and boundary reference comparisons) vs. Repeated recordings 0-2 h apart, Kors reconstruction matrix, and expert boundary annotations. Primary outcome: Short-interval repeat variation of 1,958 estimable non-circular ECG/VCG markers (median intraclass correlation coefficient [ICC(A,1)] and repeatability coefficients [RC]).

Q: What does this research mean for the field?

Automated ECG and VCG marker measurements in LoopLens 10.1 demonstrate substantial short-term repeat variation (median ICC 0.468; RC 27.27 ms for QRS and 61.78 ms for QTcF) and marked dependence on reconstruction matrix choice and fiducial boundary definitions. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Leo Probst

doi:10.21203/rs.3.rs-11410667/v1

October 20, 2026Open Access

Repeat variation and definition dependence of ECG/VCG markers: an exploratory audit of one implementation (LoopLens 10.1)

View Full Paper

Ask AI

Bookmark

Share

Key result

LoopLens ECG and VCG markers show modest repeatability with a median ICC of ~0.47.

n=10,187

Why the study?

Physiological names do not fully determine the measurement definition of an ECG or VCG marker, motivating an audit of how values in one implementation (LoopLens 10.1) vary across repeat recordings, reconstruction matrices, and boundary references.

Key Points

Evaluate how electrocardiogram (ECG) and vectorcardiogram (VCG) marker values in LoopLens 10.1 vary between routine repeat recordings, reconstruction matrices, and fiducial boundary references across a 2,005-marker catalog.
Audited 2,005 ECG and VCG markers in MIMIC-IV-ECG (15,693 repeat pairs 0–2 h apart; 5,648 recordings for matrix comparisons), LUDB for expert boundaries, and PTB-XL repeat pairs up to 24 h apart.
Tested the effect of replacing the Kors reconstruction matrix with the inverse Dower matrix and evaluated the impact of artificial ±4-ms boundary shifts.
The median ICC(A,1) of 1,958 estimable non-circular markers was 0.468 (311 reached 0.75), with repeatability coefficients of 27.27 ms for QRS duration, 61.78 ms for QTcF, and 82.41° for the QRS-T area angle.
Replacing the Kors with the inverse Dower matrix yielded 1,045 dependent, 684 conditional, and 0 stable outputs among 1,729 non-identical estimable non-circular markers, while ±4-ms boundary shifts produced 496 dependent of 1,724.
Against LUDB expert annotations, median P-onset error was +14 ms with extreme-boundary references and +1 ms with lead-median references, with mean absolute error shifting from 31.88 ms to 23.70 ms.

Study Design

Type

Observational (n=10,187)

Blinding

Question-specific non-independent analyst blinding for replication test partition

Multicenter

No

Structured PICO

P

Population

10,187 patients (15,693 ECG pairs) undergoing routine diagnostic 12-lead ECGs recorded 0 to 2 hours apart during the same admission at Beth Israel Deaconess Medical Center.

E

Exposure

LoopLens version 10.1 (an in-house Python pipeline for ECG delineation, VCG reconstruction, and calculation of a catalog of 2,005 ECG/VCG markers) evaluated across repeat recordings 0–2 h apart, reconstruction matrix substitutions (inverse Dower vs. Kors regression matrix), and boundary shifts (±4 ms).

C

Comparator

Alternative reconstruction matrices (Kors regression matrix vs. inverse Dower matrix), expert boundary references in LUDB (extreme-boundary vs. lead-median references), machine-reported ECG measurements in MIMIC-IV-ECG, BRAVEHEART 1.8.2 software outputs, and source-annotation routes in ECGRDVQ.

O

Outcome

Short-interval repeat variation (intraclass correlation coefficient ICC(A,1) and repeatability coefficient RC = 1.96 SD of paired differences), reconstruction matrix and delineator-input dependence (technical classification into stable, conditional, or dependent), and boundary errors against expert annotations.surrogate

Limitations

Properties examined are specific to a single software implementation (LoopLens 10.1) developed by the single author.
Repeat variation (RC and ICC) in routine clinical pairs includes true physiological change and cannot isolate measurement error.
Internal replication criteria for repeatability in the person-disjoint test partition were not met.
External datasets (LUDB and PTB-XL) were previously used in LoopLens development and do not provide independent validation.
LUDB amplitudes are normalized per lead (1 mV span), making boundary error findings non-transferable to physical amplitudes.
No independent human analyst blinding was performed.
Properties examined are of a single implementation (LoopLens 10.1) developed by the author and do not establish general failure of VCG markers or clinical thresholds.
ICC and repeatability coefficients in routine clinical pairs include true physiological change between recordings and selection effects, and cannot separate measurement error from true change.
The person-disjoint test partition had already been opened for other exploratory analyses, and no independent human analyst blinding was used.
LUDB and PTB-XL were used earlier in LoopLens development and calibration, so neither provides independent validation, and LUDB amplitudes are normalized per lead.
Several analyses (including route differences, proxy sensitivity, circular markers, and machine-reported measurement comparisons) were added post hoc after the first draft.

Implication

This exploratory audit of the LoopLens 10.1 ECG/VCG pipeline demonstrates substantial short-term repeat variation in routine clinical ECGs and widespread sensitivity of calculated markers to VCG reconstruction matrices and fiducial boundary definitions, highlighting that physiological marker names do not guarantee measurement stability or interchangeability.

Abstract

Objective Physiological names do not fully determine the measurement definition of an electrocardiogram (ECG) or vectorcardiogram (VCG) marker. We examined how the values of one implementation (LoopLens 10.1) change between routine recordings 0–2 h apart, between two reconstruction matrices and between boundary references, in a catalog of 2,005 markers. Approach: Development-data audits in MIMIC-IV-ECG (15,693 pairs; 5,648 recordings for the matrix comparison), expert-boundary comparison in LUDB, repeat pairs from PTB-XL and public randomized QT controls. Replication was prespecified in a person-disjoint test partition, previously opened for other analyses. Main results: The median ICC(A,1) of 1,958 estimable non-circular markers was 0.468 (311 reached 0.75); an ICC cannot separate error, change and low between-person spread. The repeatability coefficient (RC, 1.96 SD of paired differences, including true change) was 27.27 ms for QRS duration, 61.78 ms for QT corrected by the Fridericia formula (QTcF) and 82.41° for the QRS-T area angle. Under author-fixed rules, replacing the Kors with the inverse Dower matrix gave 1,045 dependent, 684 conditional and no stable outputs among 1,729 non-identical estimable non-circular ones, yet ± 4-ms boundary shifts alone gave 496 dependent of 1,724; 29 circular markers were not classified. Against LUDB experts (amplitudes normalized per lead), the median P-onset error was + 14 ms with extreme-boundary and + 1 ms with lead-median references. PTB-XL repeat variation (pairs up to 24 h apart) had equal or lower point estimates than MIMIC (R 0.77–0.99), with no class decided. In public randomized recordings, the production (standard LoopLens) and source-annotation routes differed in 7 of 16 estimable treatment contrasts (exploratory, unadjusted). In the test partition, the historical matrix classification and P availability replicated; repeatability criteria were not met, with higher test ICCs.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Leo Probst (2026) conducted an observational in Routine diagnostic ECG/VCG evaluation (n=10,187). LoopLens 10.1 ECG/VCG marker pipeline (repeat variation, Kors vs inverse Dower reconstruction matrix, and boundary reference comparisons) vs. Repeated recordings 0-2 h apart, Kors reconstruction matrix, and expert boundary annotations was evaluated on Short-interval repeat variation of 1,958 estimable non-circular ECG/VCG markers (median intraclass correlation coefficient [ICC(A,1)] and repeatability coefficients [RC]). LoopLens 10.1 ECG/VCG markers showed a median ICC of 0.468 across 15,693 pairs 0-2 h apart (RC 27.27 ms for QRS, 61.78 ms for QTcF), with 60.4% of non-circular outputs dependent on matrix choice.

synapsesocial.com/papers/6ae28f3bcbadf70daf6c4440 https://doi.org/10.21203/rs.3.rs-11410667/v1

View Full Paper

Ask AI

Bookmark

Share

Repeat variation and definition dependence of ECG/VCG markers: an exploratory audit of one implementation (LoopLens 10.1)

Key Points

Study Design

Structured PICO

Limitations

Implication

Abstract

Discussion

Cite This Study

Repeat variation and definition dependence of ECG/VCG markers: an exploratory audit of one implementation (LoopLens 10.1)

Key Points

Study Design

Structured PICO

Limitations

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading