KROOM: Knee Range of Motion Monitoring System for personalize rehabilitation in Osteoarthritis Patients. A machine learning exploratory study for more define and flexible boundaries around walking and stair ambulation clusters in free-living conditions

The KROMM, “Knee Range of Motion Monitoring” is a knee wearable electrogoniometer composed of a knee angular sensor and a three-axis accelerometer placed on the thigh. The clinical relevance of this device is guarantee to the clinicians the possibility of an accurate assessment of the knee range of motion during walking and stair ambulation and to the patients with knee osteoarthritis a personalized rehabilitation also in the home setting. The aim of this stage of the study is developing the next version of the KROMM system achieving three main goals: 1.improving the activity detection capability of the algorithm incorporated into the sensors; 2.improving the ergonomics of the design for avoid sudden breaks and uncomfortable usability; 3.improving the user- friendliness of the app on the smartphone for help the patients in the home environment giving information about the time spent in each activity. In particular, the focus of this thesis is the implementation of the activities detection mechanism of this platform that will construct more flexible and refined decision boundaries around the clusters of walking and stairs ambulation amongst dynamic, heterogeneous, and individualized free living activities. In the first session of the study 17 healthy subject, aged 18-64 years, working in the Motion Analysis Laboratory at Spaulding Rehabilitation Hospital, were selected to perform the following instructed activities: sitting, standing, walking, ascending stairs, descending stairs, riding an exercise bicycle, using a rowing machine. Walking, rowing and bicycling were performed at different speeds. These scripted data were segmented, filtered and used for the construction of the model. The features selection step was performed using the relief and the DBI index and the classification with the application of a hierarchic random forest classifier with 100 trees. The first classifier of the cascade classified between “activities of interest” and “activities of no interest”; then between the “activities of interest” the second one discriminated between “walking “, “descending stairs” and “ascending stairs”. The hypothesis is that the segmentation and the label of the instances will influence the features selection step and the classification and that the method of cross validation will affect the classification step. For verifying this hypothesis different attempts have been effectuated, changing one parameter at the time, starting from a model with the following parameters: -??Window of 5 s with no overlap -??Different labels for all the activities -??10 cross fold validation method This model was then validated on the unscripted data, data recorded for 9 hours for 4 of this 17 subjects, comparing the predictions of the classifier with the labels obtained from the images of a wearable camera that was wore facing the anterior direction of the human body. In this second part of the study to the subjects was given also a mobile phone with an application for reminding the activities to do. The aim of this thesis is providing a flexible classifier for a personalized rehabilitation a systematic approach to detect instances of ground walking and stair ambulation in uncontrolled living conditions and to understand how the borders of the clusters of the activities move depending on the number, the correlation and the labels of these instances.