Kinematic Data Augmentation Using a Spatiotemporal Dual-Discriminator Generative Adversarial Network for Joint Angle Prediction of Infant Crawling

Abstract

Abstract. Accurately generating joint motion trajectories for infant crawling is crucial for developing effective control strategies for exoskeletons. However, the limited availability of infant crawling data, owing to the high costs of data collection and privacy concerns, presents a challenge to the performance of such models and controllers. This study introduces a novel spatiotemporal dual-discriminator generative adversarial network (SDGAN) to generate synthetic kinematic data for infant crawling. The generator produces 3D joint coordinate sequences (101 time steps × 36 dimensions for 12 joints) by learning both the spatial joint relationships (via a spatial discriminator) and the temporal dynamics (via a temporal discriminator). To evaluate the model's effectiveness, the SDGAN was compared with existing benchmark models—the TimeGAN, decision-aware conditional GAN (DAT-GAN), and multivariate time series GAN (MTS-GAN). Additionally, we assessed the impact of varying synthetic-to-real data mixing ratios (0:1, 1:2, 1:1, 3:2, 2:1, 5:2, and 3:1) on the accuracy of joint angle predictions using a long short-term memory (LSTM) network. The SDGAN model significantly outperformed the benchmark models across key evaluation metrics. In terms of distribution similarity, the SDGAN achieved the lowest average Jensen–Shannon (JS) divergence (0.027), with those of TimeGAN, DAT-GAN, and MTS-GAN being 0.046, 0.080, and 0.085, respectively. The improvements were statistically significant for most joints (p < 0.05), particularly the left elbow and left knee, indicating closer alignment of the generated samples with real infant crawling data. Furthermore, incorporating SDGAN-generated data at a 1:1 synthetic-to-real mixing ratio resulted in the best joint angle prediction performance, with a mean absolute error (MAE) of 1.36 deg, representing a 40.4% reduction compared to that using only real data (MAE = 2.28 deg). This improvement was statistically significant across all major joints (p < 0.05), highlighting the practical benefits of the SDGAN for data augmentation in pediatric kinematic modeling. The results suggest that the SDGAN is a promising approach for addressing the challenge of limited infant crawling data and improving joint angle prediction accuracy in rehabilitation applications.