ECG-based approaches for predicting structural heart disease (SHD) are promising. A previous study (EchoNext, 2025) demonstrated that a 1D convolutional neural network (1D-CNN) model could predict SHD with high performance. Most conventional models process ECGs as temporally ordered waveforms. However, ECGs also implicitly encode spatial information about cardiac electrical activity. We explicitly reconstructed this spatial information as three-dimensional point clouds and extracted geometric features using a PointNet++ model. We also developed an EchoPoint model, a novel framework that integrated the PointNet++ model with the EchoNext model. EchoPoint improved AUROC, AUPRC, and F1 scores compared with EchoNext in 12-lead and single-lead settings. These findings suggest that geometric features of 3D point clouds derived from ECG signals improve SHD prediction, highlighting their potential for large-scale screening using wearable ECG devices.