HeadLite: Compact Multimodal Sensor Fusion for Stride Length Estimation From Smart Insoles

Abstract

Smart insoles provide a practical platform for ambulatory stride length estimation by combining inertial and plantar-pressure sensing. Stride length is an informative spatial gait marker linked to mobility, stability, and fall risk, and estimating it from a compact, low-power wearable is a prerequisite for continuous gait monitoring and screening outside the gait laboratory. However, existing multimodal smart-insole fusion architectures often flatten temporal features from heterogeneous channels into a single high-dimensional representation, entangling modality contributions and making low-dimensional anthropometric metadata difficult to exploit. We propose <bold>HeadLite</bold>, a compact metadata-aware multimodal fusion architecture based on branchwise temporal compression: each sensor modality is encoded independently and substantially compressed along the time axis before being fused with anthropometric metadata for stride length regression. Under strict 32-subject leave-one-subject-out (LOSO) evaluation on 7592 strides from a 46-channel smart-insole dataset collected for this study, HeadLite reduces mean absolute error (MAE) by 31.6% relative to a reproduced flatten-based prior smart-insole baseline while using 96% fewer parameters. It consistently outperforms recurrent and attention-based alternatives. The resulting error is about <inline-formula> <tex-math notation="LaTeX">${3}.{7}\%$ </tex-math></inline-formula> of the mean stride length (minimal detectable change (MDC) of <inline-formula> <tex-math notation="LaTeX">${12}.{2}$ </tex-math></inline-formula> cm), a level suited to out-of-laboratory monitoring and screening rather than laboratory-grade measurement; even a single deployable network (<inline-formula> <tex-math notation="LaTeX">${1}.{73}$ </tex-math></inline-formula>M parameters, submillisecond GPU inference) reduces error by roughly <inline-formula> <tex-math notation="LaTeX">${21}\%$ </tex-math></inline-formula> relative to the reproduced baseline, so the improvement does not rely on ensembling. A direct head-to-head ablation shows that anthropometric metadata meaningfully improves accuracy under HeadLite but provides no statistically significant benefit under the same flatten-based fusion of identical sensor inputs. This finding supports fusion-level dimensional balance as an important architectural condition for effective metadata utilization in subject-independent smart-insole stride length estimation.