A Machine-Learning Classification of Walking in the Community using Inertial Sensors and Smart Insole

Wednesday, November 9, 2022

4:04 PM – 4:09 PM

Location: Station 2

Presenting Author(s)

Hogene Kim, PhD

Senior Research Officer
National Rehabilitation Center
Gangbuk-gu, Seoul-t'ukpyolsi, Republic of Korea
KS

Kyeong-Jun Seo, MS

Researcher
NRC
Gangbuk-gu, Seoul-t'ukpyolsi, United States
Ji-Eun Cho, PhD

Researcher
National Rehabilitation Center
Gangbuk-gu, Republic of Korea

Research Objectives: - To improve the success rate to classify five different common walking conditions using various machine learning schemes

Design: - Cross-sectional study

Setting: - Walking Conditions during Community Walking; 1) Even surface; 2) Slope up; 3) Slope down; 4) Stair up; 5) Stair down

Participants: Five healthy participants ( 37.8+/-7.25 yrs, F:4)

Interventions: - Walking over five different kinds of gait activities in community 1) walking on even surface; 2) Slope up; 3) Slope down; 4) Stair up; 5) Stair down

- Analyzed post-processed data from sensors to classify the walking condition using machine learning from the signals of five Inertial Measurements Units (IMU) , attached to waist, bilateral lower and upper legs, and smart insoles that measured total and hind/forefoot pressure distribution

Main Outcome Measures: - Analyzed 32934 gait data out of 47,033 training data using an Artificial Neural Networks (ANN), which was a Feed Forward Neural Network (FFNN)
- Investigated 3 different initial sets of variables; 1) Validation ratio 10%, Test ratio 20%, 2) Validation data ratio 15%, Test data ratio 15%, 3) Validation ratio 20%, Test ratio 10%

Results: - Highest accuracy was achieved when validation ratio 15% and test data ratio 15% with the hidden layers was 1000 using FFNN (98.3%)

Conclusions: - 5 gait conditions were successfully categorized using FFNN with higher accuracy rate
- Machine learning and senor network would be potentially useful to assist pathological gait, stroke, in the disabled by identifying walking condition in the community

Author(s) Disclosures: All authors declare that no conflicts or lack thereof on this presentation.

Learning Objectives:

Upon completion, participant will be able to describe that the machine learning using signals from inertial measurement units in lower extremities and foot pressure provided higher success accuracy to detect walking conditions in community environments.
Upon completion, the highest accuracy was achieved when validation ratio 15% and test data ratio 15% with the hidden layers was 1000 using feed forward neural network to detect walking conditions in community environments.
3. Upon completion, the real-time classification using machine learning is a promising method to predict pathological gait patterns in community environments, e.g. patients with stroke hemiparesis.