Exploring the use of Mitsui’s flexible optical sensor as a means to monitor knee flexion-extension angular displacement during ambulation

A branch of the wearable sensor research focuses on monitoring the human body joints movements to prevent injuries due to musculoskeletal disorders. Different types of sensors and technologies are used as a core of a smart joint monitoring system, one of these are optical fiber sensors. Optical fiber sensors are used to monitor joint angles with the main advantages to have a high resolution, a light-weight and to be flexible. For these reasons, they can be integrated into stretchable skin-tight fabrics to create a wearable joint monitoring system. In this thesis, a prototype developed by Mitsui Chemicals has been tested. The prototype is composed of a sleeve with 5 integrated optical fiber sensors placed medially, laterally, inside-medially and inside-laterally than the knee and above the patella. The purpose of this work is to evaluate the performance of the Mitsui's optical fiber sensors in predicting knee flexion-extension angular displacement. The first step of the study is the creation of a dataset to train a machine learning algorithm. Twenty healthy participants have been selected to perform the data collection. They have been requested to wear the prototype during ambulation on a treadmill at different speeds while the voltage outputs and the knee angles have been recorded using the Vicon Motion Capture system. Three sizes of the prototype have been used during the data collection: for each subject, the size with a closer fit that won't create discomfort to the subject has been chosen. An outlier detection algorithm has been developed to analyze the voltage outputs to find possible outliers due to a not optimal fit of the sleeve. This algorithm recognizes as outliers signals with a low correlation value and large amplitude respect the other signals of the sensor with the same position in the sleeve prototype. The outliers are subsequently removed from the dataset by a custom replacement algorithm. Demonstrating the high correlation among medially, laterally and patella sensors, and inside-medially and inside-laterally sensors, it's possible to replace outliers of one sensor with the data of the same correlated sensor. Two different types of models to predict the knee angle are used: one model without memory, the Random forest (RF), and one model with memory, the Long short-term memory (LSTM). The models are applied to the dataset with a leave-one-subject-out (LOSO) cross-validation (CV) approach and their performances are evaluated in terms of root-mean-squared error (RMSE). The overall RMSE obtained for the RF model is 5.2° and the LSTM model is 5°. The average error is similar, however, the LSTM performs better than the Random Forest in 80% of the subjects.