Abstract <p>Patients with lumbar disc herniation exhibit distinctive gait patterns, making gait-based analysis a promising approach for disease detection. Gait data are periodic multivariate time series (MTS), where discriminative information is often distributed across multiple temporal scales. However, existing methods primarily focus on intra-sequence temporal features and overlook inter-sequence correlations and multi-period characteristics. To address these issues, we propose a Temporal and Correlated features extraction Network (TCGNet). Fast Fourier Transform (FFT) is used to identify dominant frequencies and decompose MTS into multi-period subsequences. A dual-branch architecture extracts intra-sequence temporal features via convolution and inter-sequence correlated features via adaptive graph convolution, followed by feature fusion. Experiments on nine benchmark datasets demonstrate superior classification performance, and gait-based lumbar disc herniation diagnosis further validates the effectiveness of TCGNet.</p> Graphic abstract <p></p>