medRxivSource publication:
Loading content…
Cardiac MRI reading no longer rests only on ejection fraction and other predefined measurements: a healthy cardiac digital twin generated from each person's age, sex and body habitus yields a deviation score that still improves atrial fibrillation risk stratification among people with preserved ejection fraction.
Cardiac imaging gains a channel alongside conventional measurements: each person's cardiac dynamics are compared with a healthy reference generated from their own age, sex and body habitus, and the deviation itself becomes a continuously scored phenotype. Clinical practice had used a small set of predefined measurements, ejection fraction, ventricular volumes, myocardial mass and strain, which compress three-dimensional, time-varying cardiac dynamics into a few scalars, and patients with similar ejection fractions can follow markedly different clinical trajectories. In the HeartSSM cohort of 52,365 participants, the deviation score retained incremental value for atrial fibrillation after adjusting for ejection fraction, ventricular volumes, myocardial mass and strain, with an incremental C-index of +0.049 (5-fold cross-validated +0.055; likelihood-ratio P = 8.8×10⁻⁹), against a Gaussian-noise placebo control of +0.0004, P = 0.77.
The approach first builds a conditional healthy reference inside a generative cardiac representation space, then measures the distance between an individual and that reference. The HeartSSM cohort uses a 128-dimensional four-chamber statistical shape descriptor, with polynomial ridge regression (degree 2, regularisation α = 1.0) trained on 7,551 participants free of cardiovascular disease to learn the expected shape descriptor from age, sex, weight, height, blood pressure, smoking, alcohol use and physical activity; the MeshHeart cVAE cohort instead takes the pretrained conditional decoder's expected mesh sequence at the prior's mode z = 0 conditioned on the subject's own demographics, with no additional fitting. In the MeshHeart cVAE cohort, mean deviation was 23.2% lower among healthy participants (n = 21,289) than among those with a recorded cardiovascular risk factor or diagnosis (n = 10,648; Mann–Whitney P ≈ 0); in held-out validation of the HeartSSM model, empirical coverage of the nominal 95% prediction interval was 0.946.
The deviation score separates risk inside the preserved-ejection-fraction population, where conventional measurements give little. Preserved ejection fraction (≥50%) has been the range where conventional biomarkers carry limited prognostic information, and in the HeartSSM preserved-EF subgroup (n = 48,976) the deviation score's incremental value for atrial fibrillation was larger than in the unstratified cohort. That subgroup had 102 atrial fibrillation events, with incremental C-index +0.068 (5-fold cross-validated +0.063; likelihood-ratio P = 3.3×10⁻⁸; Gaussian-noise placebo −0.001, P = 0.58); the independent MeshHeart cVAE preserved-EF subgroup (n = 30,028) gave +0.024 for stroke (cross-validated +0.025, P = 2.6×10⁻⁴), and the two representations did not reproduce each other for any single outcome.
The work addresses cardiac magnetic resonance imaging research and cardiovascular risk stratification, in UK Biobank participants imaged with a standardised 1.5T protocol who have shape descriptors or mesh sequences, plus the GenScan/Hammersmith healthy volunteers used for the cross-site check. The authors' stated next steps are prospective validation of preserved-EF risk stratification in independent clinical cohorts and longitudinal trajectory phenotyping that tracks how individual deviation profiles evolve with repeat cardiac imaging. The authors also list extending the same normative-twin principle to other organ systems as a long-term direction rather than a focus of this study.
The UK Biobank cohort is predominantly of European ancestry and healthier than average, and the external validation cohort is itself a small healthy-volunteer sample from the same broad UK population, so validation in larger, ethnically diverse and clinically enriched cohorts remains essential. The normative model cannot capture all sources of physiological variation, and some deviations may reflect unmeasured confounders rather than disease-relevant change. Only the MeshHeart cVAE representation is directly generative; the HeartSSM descriptor allows deviation phenotypes to be interpreted as shape modes but not decoded into three-dimensional cardiac motion. In a later sensitivity analysis in the preserved-EF subgroup, adding maximum left-atrial volume and left-atrial ejection fraction removed the reliable incremental value of the HeartSSM score for atrial fibrillation (cross-validated ΔC = +0.004, P = 0.14), so the authors do not claim information beyond atrial measures. Follow-up ended on 2021-01-11, 59% of preserved-EF atrial fibrillation events occurred within two years of imaging, analyses were not corrected for multiple testing, and event counts were limited in some subgroups.