LAUSR

Walking speed (WS), the “sixth vital sign,” is critical for health assessment. Its estimation using only insole pressure sensors (IPSs), however, remains a challenge. We introduce a deep-learning framework combining multimodal ResNet-50 with multioutput regression (MOR) for estimating WS, stride length, and step length. In experiments with 25 healthy participants walking at self-selected speeds, the method achieved high accuracy root mean squared error (RMSE) RMSE: 3.93 cm/s, adjusted ${R}^{{2}}$ : 0.89), outperforming previous studies. Ablation studies demonstrated that removing either the multimodal architecture or MOR reduced the performance significantly, underscoring the critical role of both components in enhancing the model’s accuracy. Feature importance analysis revealed that mean midfoot center of pressure (CoP) velocity and second double support (SDS) time are pivotal to WS estimation, while participant-specific variables (shoe size, height) also contribute substantially. Key factors included mean midfoot CoP velocity and SDS time, highlighting the potential for integrating contralateral data and additional biomechanical parameters (e.g., leg length) in future refinements. Overall, these findings demonstrate that WS can be robustly estimated using only IPS, avoiding the need for extra sensors.