Predicting five-year mortality in patients with left ventricular hypertrophy based on dynamic radiomics from transthoracic echocardiography

Background

Left ventricular hypertrophy (LVH) is a critical structural adaptation associated with an increased risk of adverse cardiovascular events and all-cause mortality. Within this population, hypertrophic phenocopies, comprising diverse etiologies, require precise differentiation due to the availability of disease-specific therapies. However, hypertrophic phenocopies frequently exhibit aggressive progression and poorer prognoses in the absence of adequate treatment. Transthoracic echocardiography (TTE) is the most available cardiac imaging tool, and Radiomics is powerful to extract high-dimensional quantitative features from medical images. Therefore, precise survival prediction from TTE is crucial for optimizing risk stratification and clinical management in patients with LVH.

Purpose

We aimed to develop a dynamic radiomics-based deep learning model derived from TTE for predicting 5-year all-cause mortality in patients with LVH.

Methods

We conducted a retrospective cohort study between 2010 and 2025 in Taiwan. The inclusion criteria were defined as a left ventricular (LV) mass index >95g/m² in women or >115g/m² in men, or a maximal LV wall thickness ≥13mm on TTE. The endpoint of 5-year all-cause mortality was the duration between the TTE study date and the date of death. The workflow, illustrated in Fig.1, included the following: (1) interventricular septum (IVS) segmentation, (2) dynamic radiomics process, and (3) mortality prediction.

One TTE study was scanned per patient, with each study comprising multiple cardiac cycles. We contoured IVS in apical four and apical five chamber views and subdivided IVS into upper, lower, border, and peripheral segments. A total of 310 dynamic radiomic features were extracted (62 features per segment). Following feature preprocessing, the top 10% of features ranked by SHAP values were utilized to train a temporal convolution network for mortality prediction.

Results

The study cohort comprised 275 patients (61.8±12.4 years, 61.5% male), including 109 with hypertrophic cardiomyopathy (HCM), 103 with Fabry disease (FD), and 63 with cardiac amyloidosis (CA). The 5-year all-cause mortality rate was 8.7%, with disease-specific rates of 4.6% in HCM, 10.7% in FD, and 12.7% in CA. The median follow-up duration was 2.989 years for surviving patients. We achieved an accuracy of 96.3%, an F1-score of 0.667, and an AUC of 0.868 for predicting 5-year mortality at the patient level (Fig.2A). SHAP values, evaluating the contribution to model decision-making, showed that the systolic phase was more important (Fig.2B). The four subdivided IVS segments contributed equally with similar SHAP values (Fig.2C).

Conclusion

We successfully developed a dynamic radiomics-based model to predict 5-year all-cause mortality for patients with LVH. The findings demonstrate the potential of TTE-derived radiomics to assist clinicians in identifying high-risk populations and implementing precision management strategies.Workflow of This StudyModel Performance of Mortality Prediction