Developing a multi-modal sensor network to detect and monitor conditions consequential to knee injuries

Degenerative joint diseases are major causes of mobility problems in older people. These conditions are not curable, but early detection can allow treatment and lifestyle changes which can alleviate the symptoms and slow progression. The objective of this study was to develop and characterise a new system for the detection, monitoring and classification of acoustic emissions from knee joints powered by deep learning, with these being considered as potential aids to diagnosis of conditions such as osteoarthritis. Two studies were conducted on 30 and 40 adult volunteer participants of both sexes and of varied age. In the first study the participants were required to perform six sets of three sit-stand-sit cycles (three for each leg that was monitored) on 5 different occasions. In the second study the volunteers performed three sets of five sit-stand-sit cycles at different tempos with the aid of a visual metronome. Twenty-seven retroreflective markers were placed on each participant at specific body landmarks recorded by nine Oqus 700+ cameras operating at 200Hz to monitor movement. In the first study the exercise was performed with one foot on a force platform recording at 2000Hz. A sensitive condenser microphone with a wide frequency response was connected to a Laryngograph DSP Unit recording at 16000 Hz. Results were obtained from all 30 participants on five different days within a period of three weeks specific to each participant to study the reproducibility of the system and to avoid structural changes to participants’ knees over a prolonged period of time. A MATLAB code was developed to process the data and generate a collage of visual representations of acoustic emission information combined with data from kinematic variables, which was then analysed using classical and custom processing techniques and latterly used for training convolutional neural networks. Due to a relatively small sample size for training of neural networks a novel method for data augmentation was developed. The network was consequently used to classify the images into three age categories with 89.5% accuracy. Results provide clear acoustic signals showing a distinctive sequence of impulse-decay forms occurring naturally during each sit-stand-sit cycle. There are distinct differences between the acoustic signals emitted from younger healthy knees and those from aged knees. The second study was inspired by the results of cross correlations between acoustic emissions of the knee joints and kinematic variables such as knee angle, knee angular velocity and angular acceleration. A clear periodicity indicated a strong dependence of acoustic emission on angular velocity and acceleration. 40 participants of varied age and both sexes took part in this study. They were asked to perform five sit-stand-sit cycles matching three different tempos (120, 180 and 240 beats per minute). Once again, the augmentation method was employed, and the convolutional neural networks were trained. This time the data was classified into three different tempo categories. The network classified the images with 81.6% testing accuracy. The results also clearly indicate that with increased knee angular velocity the knee generates more acoustic events occurring at the same location as at the lower velocities. This work demonstrates that the system could be used as an indication of the state of health of a human knee during movement. Further research involving clinical populations with existing arthritic conditions should focus on validating the integration of acoustic emission with kinematic variables and neural networks as a biomarker for knee degeneration, emphasising larger datasets and cross-validation with clinical and imaging assessments.