Gaussian Process-driven Hidden Markov Models for Early Diagnosis of Infant Gait Anomalies

Abstract

Gait analysis is critical in the early detection and intervention of motor neurological disorders in infants. Despite its importance, traditional methods often struggle to model the high variability and rapid developmental changes inherent to infant gait. To address these challenges, we propose a probabilistic Gaussian Process (GP)-driven Hidden Markov Model (HMM) to capture the complex temporal dynamics of infant gait cycles and enable automatic recognition of gait anomalies. We use a Multi-Output GP (MoGP) framework to model interdependencies between multiple gait signals, with a composite kernel designed to account for smooth, non-smooth, and periodic behaviors exhibited in gait cycles. The HMM segments gait phases into normal and abnormal states, facilitating the precise identification of pathological movement patterns in stance and swing phases. The proposed model is trained and assessed using a dataset of infants with and without motor neurological disorders via leave-one-subject-out cross-validation. Results demonstrate that the MoGP outperforms Long Short-Term Memory (LSTM) based neural networks in modeling gait dynamics, offering improved accuracy, variance explanation, and temporal alignment. Further, the predictive performance of MoGP provides a principled framework for uncertainty quantification, allowing confidence estimation in gait trajectory predictions. Additionally, the HMM enhances interpretability by explicitly modeling gait phase transitions, improving the detection of subtle anomalies across multiple gait cycles. These findings highlight the MoGP-HMM framework as a robust automatic gait analysis tool, allowing early diagnosis and intervention strategies for infants with neurological motor disorders.