AI-assisted longitudinal cardiac phenotyping identifies domain-specific remodeling patterns in Fabry cardiomyopathy

Cardiac involvement in Fabry disease spans structural hypertrophy, mechanical dysfunction, and electrical conduction abnormalities that rarely progress in parallel. Although such discordant longitudinal behavior is commonly encountered in clinical practice, it has not been systematically organized within a coherent longitudinal framework.

We conducted a retrospective multicenter longitudinal study of 38 patients managed across three geographically distinct affiliated hospitals within the MacKay Memorial Hospital system, located in Taipei, Tamsui, and Hsinchu. Three major domain-specific remodeling patterns, including structural, mechanical, and electrical domains, were assessed and indexed by left ventricular mass index (LVMi), left ventricular global longitudinal strain (LVGLS), and QRS duration using standard 12-lead electrocardiography (ECG), respectively. All echocardiographic domain measures were conducted using an AI-assisted deep learning-based platform (Us2.ai), which enables standardized longitudinal processing of multi-parametric measurements. Domain-specific longitudinal within-patient changes were classified by the sign of the numerical change (positive vs. negative), with adverse or non-adverse designation defined according to the domain-specific clinical meaning of each parameter, and compared between patients with and without enzyme replacement therapy (ERT).

During a median of 5.3 years (interquartile range, 2.8-7.1 years) follow-up, electrical remodeling showed an adverse trend, with QRS duration increasing from 128.21 ± 31.18 ms to 139.63 ± 37.95 ms (Δ + 11.42 ms, 95% CI 5.93-16.92; p < 0.001). Mechanical remodeling assessed by LVGLS changed from -14.74 ± 5.83% to -16.69 ± 4.44% (Δ - 1.95 percentage points; 95% CI, -4.05 to 0.14; p = 0.067), whereas structural remodeling by LVMi showed no significant longitudinal change (p > 0.05). In the LVGLS analysis, concordant adverse remodeling was most frequent (18/38, 47.4%), followed by electrical-leading adverse remodeling (12/38, 31.6%), mechanical-leading adverse remodeling (5/38, 13.2%), and concordant favorable remodeling (3/38, 7.9%). The phenotype distribution differed significantly between untreated and ERT-treated patients (p < 0.001).

AI-assisted longitudinal monitoring enables consistent characterization of non-parallel, domain-specific change patterns. Electrical-mechanical phenotypes were heterogeneous, with a significant difference in phenotype distribution between untreated and ERT-treated patients.