Autoencoder-driven stride length estimation for individuals with Parkinson's disease using inertial measurement...

Gait disorders in Parkinson's disease (PD) are often characterized by reduced stride length, particularly under off-medication conditions. Inertial measurement units (IMUs) embedded in footwear provide a practical and environment-independent method for assessing stride length. Although deep learning approaches have recently shown promise for stride length estimation from inertial signals, most models have been trained on healthy participants, limiting their applicability to populations with gait impairments.

We developed a deep learning model to predict stride length in 10 individuals with PD using tri-axial accelerations and angular velocities recorded from footwear-mounted IMUs. The model incorporated an autoencoder for feature extraction, followed by a fully connected regression network for stride length prediction.

Leave-one-out cross-validation against ground-truth stride lengths from a GAITRite walkway demonstrated strong agreement, with an R² of 0.8619 and limits of agreement of ±18.11 cm for 1-second windows preceding heel strike. To assess robustness against stride segmentation errors, simulated heel-strike timing offsets of up to ±0.5 s were introduced. Although performance declined with increasing offsets, the model maintained an R² of 0.8033 and limits of agreement of ±21.58 cm at ±0.5 s, indicating that explicit stride segmentation is not required.

These findings demonstrate the feasibility of stride length estimation in PD without explicit stride identification, paving the way for unobtrusive gait monitoring and objective evaluation of disease progression and therapeutic interventions.