Reconstruction of gait variables from 3D foot geometry and sparse plantar pressure landmarks using deep learning.

This study quantified the ability of static three-dimensional foot geometry and sparse plantar pressure landmarks to reconstruct stance-phase plantar pressure distributions. Thirty-five healthy adults underwent 3D plantar scanning and dynamic pressure measurements during level walking. Two landmark configurations were defined: 16 sensors covering primary load-bearing zones and a reduced 2-sensor core set at the heel and third metatarsal head. A convolutional block attention module U-Net model was trained using 5-fold cross-validation to map the 3D foot image and landmark mask to pressure maps. Model accuracy was assessed by comparing predicted and measured vertical ground reaction force, peak plantar pressure, and centre of pressure trajectories. The 16-landmark configuration achieved 97.1 % vertical ground reaction force accuracy, 98.2 % peak-pressure accuracy, and 0.8 cm mean centre of pressure error; the 2-landmark configuration achieved 95.5 %, 98.9 %, and 1.3 cm, respectively. Errors were confined to regions within approximately 4 × 4 cm of each landmark. The results indicate that foot geometry combined with sparse anatomical landmarks provides sufficient boundary information to reconstruct biomechanically consistent pressure fields, enabling reduced-sensor insole designs for gait analysis and orthotic assessment.