MeshHeart: A Personalized 3D+t Mesh Generative Model for Unveiling Normal Heart Dynamics

Official implementation of the Nature Machine Intelligence 2025 paper

Qiao M., McGurk K.A., Wang S., Matthews P., O'Regan D., Bai W.

Overview

MeshHeart is a geometric deep generative model that reconstructs and generates personalized 3D+t cardiac mesh sequences from demographic and anthropometric conditions (age, sex, weight, height). It models cardiac motion dynamics using a graph-based latent VAE and temporal transformer, enabling normative modelling and disease deviation quantification across large-scale population data.

MeshHeart generates temporally smooth, anatomically realistic 4D cardiac meshes conditioned on age, sex, weight, and height.

Features

• 🫀 Models personalized 4D cardiac mesh dynamics from clinical attributes
• 🧠 Latent transformer VAE architecture for spatial-temporal modelling
• ⚙️ Latent delta to quantify deviation from normal heart function
• 🧬 PheWAS-enabled discovery of genotype–phenotype–disease links
• 📊 Trained on 38,309 UK Biobank subjects

Installation

```bash git clone https://github.com/your-username/MeshHeart.git cd MeshHeart conda env create -f environment.yml conda activate meshheart