Using Wearable Technologies and Machine Learning methods to Estimate Clinical Outcomes for Acquired Brain Injury Rehabilitation

The development of personalized rehabilitation strategies for patients with hemiparesis is fundamental to achieve the most effective outcome from the treatments. Clinicians are fully aware of the fact that the patients’ responsiveness to an intervention is extremely subjective and that the need to quantitatively track their motor-gains is evident. Wearable sensing technology can meet this demand through cost-effective and flexible solutions, enabling accurate assessments of movement quality and motor impairment. The approach proposed in this thesis relies on machine learning-based algorithms to estimate clinical scores through the analysis of wearable accelerometers data collected during the performance of Activities of Daily Living (ADLs). The purpose of the study is to build predictive models able to mimic the evaluation criteria currently used by clinicians, in order to define the recovery trajectory of Stroke survivors and Traumatic Brain Injury patients. Among the numerous assessment scales developed in the past years, in this project the upper limb Fugl-Meyer assessment (FMA) scale was used to quantify the severity of motor impairments, and the Functional Ability Scale (FAS) was used to evaluate the quality of movement. 3-axial accelerometers data are preprocessed with Digital Signal Processing (DSP) methods, such as segmentation and filtering, and subsequently analyzed with the aim of extracting informative features capable of defining movements properties of the study participants. Through a feature selection process, only the relevant characteristics are kept with the intention of discarding noisy and redundant data. The estimation of the clinical scores is done training and validating a regressive model using a Random Forest algorithm. Finally, the regression equation, relating the actual scores provided by the clinician and the predicted scores, is derived. In order to assess the accuracy of the algorithms, two regression problems evaluation metrics are computed: the root-mean-square error (RMSE) and the coefficient of determination (R2). The results show that this approach is efficient and that, through the Activity of Daily Living tasks, it is concretely possible to estimate patients’ rehabilitation outcomes in terms of both the movement quality and the motor impairment. The model performance is in line with the standards as R2 of 0.83 and 0.78 are reached respectively for FAS and FMA. This is solely achieved through the analysis of wearable accelerometers data, while slightly better results can be probably obtained adding clinical and demographic information about patients. This fruitful project paves the way for further studies that will presumably focus on reducing the number of sensors needed for the motor assessment and on moving the recordings from a clinical setting towards a home-based scenario.