Artificial neural network based ankle joint angle estimation using instrumented foot insoles

Current trends for long term gait monitoring relies on estimations made via machine learning. As such, this work investigates the viability of feedforward neural network (FFNN) to estimate ankle angles using ground reaction forces (GRFs) acquired from a wearable foot insole system. Inputs from nine salient gait events were selected for network training. These nine gait events are loading response (LR), pre-initial single support (pre-ISS), initial single support (ISS), post-initial single support (post-ISS), mid single support (MSS), pre-terminal single support (pre-TSS), terminal single support (TSS), post-terminal single support (post-TSS) and pre swing (PSW). Ankle angles are estimated with ρ¯>0.95 and NRMSE¯: 5.475 ± 1.34% for left leg in-sample estimations, 5.614 ± 1.1% for right leg in-sample estimations, 5.745 ± 1.642% for left leg out-sample estimations and 6.536 ± 0.9798% for right leg out-sample estimations. This method potentially eliminates the need of multiple wearable sensors and allow ankle angle estimation for long term basis with the aid of a simpler sensor layout. Therefore, the proposed work investigates the feasibility of using ANN for lower limb angle estimations from ground reaction forces measured using wearable foot insoles.