Interpretable ECG manifolds link longitudinal electrical phenotypes to heart-failure risk in hypertrophic cardiomyopathy

Q: What are the key findings of this study?

Variational autoencoder ECG representations across 650 HCM patients (2,594 visits) identified a time-dependent lateral/frontal QRS-force coordinate associated with subsequent heart-failure risk.

Q: What is the clinical evidence from this study?

Study design: Cohort. Population: Hypertrophic cardiomyopathy (HCM) (n=650). Intervention: Variational autoencoder (VAE)-derived longitudinal ECG electrical phenotypes (specifically lateral/frontal QRS-force coordinate z06_HCM) vs. Normal physiological ECG manifold (PTB-XL) and higher z06_HCM coordinate values. Primary outcome: Longitudinal drift in ECG latent coordinates and subsequent SHaRe heart-failure composite (LVEF <35%, NYHA class III/IV, cardiac transplantation, or LVAD implantation).

Q: What does this research mean for the field?

In patients with hypertrophic cardiomyopathy, a longitudinal lateral/frontal QRS-force coordinate derived from unsupervised ECG representations undergoes time-dependent drift and its lower values independently predict subsequent heart-failure risk. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

Pietro Cerveri; Marion Taconné; Valentina Corino; Luca Tommaso Mainardi

doi:10.64898/2026.11.02.26367092

November 4, 2026Open Access

Interpretable ECG manifolds link longitudinal electrical phenotypes to heart-failure risk in hypertrophic cardiomyopathy

View Full Paper

Ask AI

Bookmark

Share

Key result

Latent ECG coordinate drift links to subsequent heart-failure risk in hypertrophic cardiomyopathy.

n=650

Why the study?

The continuous organization, longitudinal evolution, and clinical relevance of heterogeneous electrical phenotypes in hypertrophic cardiomyopathy remain poorly defined.

Do variational-autoencoder ECG representations identify longitudinal electrical trajectories and predict heart-failure risk in individuals with hypertrophic cardiomyopathy?

Key Points

To develop unsupervised variational-autoencoder representations of ECG morphology to characterize continuous electrical phenotypes, longitudinal drift, and clinical risk in hypertrophic cardiomyopathy.
Trained complementary variational autoencoders on 5,194 normal-reference ECGs and 2,594 longitudinal visits from 650 individuals with hypertrophic cardiomyopathy from the SHaRe registry.
Tracked longitudinal electrical trajectories across fixed latent manifolds and used a time-conditioned change autoencoder alongside mixed-effects models to quantify within-subject phenotypic drift.
Evaluated associations between phenotypic coordinates and subsequent heart-failure risk after censoring for septal reduction therapy, atrial fibrillation, and pacing, as well as sarcomeric genotype.
Hypertrophic cardiomyopathy ECGs overlapped the physiological manifold while extending toward its periphery, with an HCM-oriented manifold reorganizing shared electrical modes.
A lateral/frontal QRS-force coordinate demonstrated persistent time-dependent drift after censoring septal reduction therapy, atrial fibrillation, and pacing, with lower values associated with heart-failure risk.
A time-conditioned change model identified four interpretable within-subject modes, while sarcomeric genotype associated with baseline phenotype position rather than trajectory rate.

Study Design

Type

Cohort (n=650)

Multicenter

Yes

Structured PICO

Do variational-autoencoder ECG representations identify longitudinal electrical trajectories and predict heart-failure risk in individuals with hypertrophic cardiomyopathy?

P

Population

650 individuals with hypertrophic cardiomyopathy (contributing 2,594 longitudinal ECG visits) from the SHaRe registry, evaluated alongside 5,194 normal-reference ECGs from PTB-XL and 335 external validation subjects from MIMIC-IV.

E

Exposure

Complementary variational autoencoder (VAE) representations (Normal-anchored β-VAE, HCM-oriented static β-VAE, and longitudinal time-conditioned Delta-VAE) applied to 12-lead median-beat ECG waveforms.

C

Comparator

5,194 normal-reference ECGs from the PTB-XL dataset (for manifold anchoring).

O

Outcome

28 primary longitudinal electrical trajectory variables (16 Normal-anchored latent coordinates, 8 HCM-oriented latent coordinates, Mahalanobis distances dN and dHCM, and 2 model-specific physical reconstruction errors) evaluated via linear mixed-effects models; and the SHaRe heart-failure composite (first occurrence of LVEF <35%, NYHA functional class III/IV symptoms, cardiac transplantation, or left ventricular assist-device implantation) for secondary clinical-risk analysis.composite

Limitations

Normal and HCM cohorts originated from different source datasets, preventing diagnostic validation of reconstruction differences.
Latent morphology coordinates were derived from median beats and do not capture intrinsic heart-rate or rhythm representations.
Limited number of incident events for the ventricular-arrhythmic composite and post-septal reduction therapy landmark analyses.
Patient-specific BLUP slopes are shrinkage estimates and not statistically independent from Delta-VAE summaries.
Normal and HCM data originated from different source datasets, so HCM-versus-Normal reconstruction differences were not interpreted as diagnostic validation.
Because the two autoencoders differed in training population and model capacity, relative reconstruction burden (∆RE) was interpreted descriptively and not as a calibrated measure of disease specificity.
Ventricular-arrhythmic composite analysis was exploratory due to the limited number of incident events.
Post-septal reduction therapy (SRT) landmark analyses were exploratory because of the small number of eligible subjects and subsequent events.
Integration of Delta-VAE summaries and trajectory slopes are not considered statistically independent validation because they are derived from overlapping serial ECG observations.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Variational autoencoder-derived ECG manifolds capture continuous morphological heterogeneity and longitudinal electrical drift in hypertrophic cardiomyopathy, identifying a specific QRS-force coordinate associated with subsequent heart-failure risk.

Abstract

Hypertrophic cardiomyopathy (HCM) produces heterogeneous electrical phenotypes whose continuous organization, longitudinal evolution and clinical relevance remain poorly defined. We developed complementary variational-autoencoder representations to map HCM ECG morphology relative to physiological variation and characterize within-subject change. Using 5,194 normal-reference ECGs and 2,594 longitudinal visits from 650 individuals with HCM, we found that HCM overlapped the physiological manifold but extended towards its periphery, while an HCM-oriented manifold largely reorganized shared electrical modes. Mixed-effects models identified longitudinal drift in selected phenotypic axes. A lateral/frontal QRS-force coordinate remained time-dependent after censoring septal reduction therapy, atrial fibrillation and pacing, and lower values were associated with subsequent heart-failure risk in baseline and time-updated analyses. A time-conditioned change VAE identified four interpretable within-subject modes aligned with longer-term trajectories. Sarcomeric genotype was associated with phenotype position but not trajectory rate.

Discussion

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

Cite This Study

Cerveri et al. (2026) conducted a cohort in Hypertrophic cardiomyopathy (HCM) (n=650). Variational autoencoder (VAE)-derived longitudinal ECG electrical phenotypes (specifically lateral/frontal QRS-force coordinate z06_HCM) vs. Normal physiological ECG manifold (PTB-XL) and higher z06_HCM coordinate values was evaluated on Longitudinal drift in ECG latent coordinates and subsequent SHaRe heart-failure composite (LVEF <35%, NYHA class III/IV, cardiac transplantation, or LVAD implantation). Variational autoencoder ECG representations across 650 HCM patients (2,594 visits) identified a time-dependent lateral/frontal QRS-force coordinate associated with subsequent heart-failure risk.

synapsesocial.com/papers/6aee6bd871a7594806220c2a https://doi.org/10.64898/2026.11.02.26367092

View Full Paper

Ask AI

Bookmark

Share

Interpretable ECG manifolds link longitudinal electrical phenotypes to heart-failure risk in hypertrophic cardiomyopathy

Key Points

Study Design

Structured PICO

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Interpretable ECG manifolds link longitudinal electrical phenotypes to heart-failure risk in hypertrophic cardiomyopathy

Key Points

Study Design

Structured PICO

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading