Sensor-based artificial intelligence to support people with cognitive and physical disorders

A substantial portion of the world's population deals with disability. Many disabled people do not have equal access to healthcare, education, and employment opportunities, do not receive specific disability-related services, and experience exclusion from everyday life activities. One way to face these issues is through the use of healthcare technologies. Unfortunately, there is a large amount of diverse and heterogeneous disabilities, which require ad-hoc and personalized solutions. Moreover, the design and implementation of effective and efficient technologies is a complex and expensive process involving challenging issues, including usability and acceptability. The work presented in this thesis aims to improve the current state of technologies available to support people with disorders affecting the mind or the motor system by proposing the use of sensors coupled with signal processing methods and artificial intelligence algorithms. The first part of the thesis focused on mental state monitoring. We investigated the application of a low-cost portable electroencephalography sensor and supervised learning methods to evaluate a person's attention. Indeed, the analysis of attention has several purposes, including the diagnosis and rehabilitation of children with attention-deficit/hyperactivity disorder. A novel dataset was collected from volunteers during an image annotation task, and used for the experimental evaluation using different machine learning techniques. Then, in the second part of the thesis, we focused on addressing limitations related to motor disability. We introduced the use of graph neural networks to process high-density electromyography data for upper limbs amputees’ movement/grasping intention recognition for enabling the use of robotic prostheses. High-density electromyography sensors can simultaneously acquire electromyography signals from different parts of the muscle, providing a large amount of spatio-temporal information that needs to be properly exploited to improve recognition accuracy. The investigation of the approach was conducted using a recent real-world dataset consisting of electromyography signals collected from 20 volunteers while performing 65 different gestures. In the final part of the thesis, we developed a prototype of a versatile interactive system that can be useful to people with different types of disabilities. The system can maintain a food diary for frail people with nutrition problems, such as people with neurocognitive diseases or frail elderly people, which may have difficulties due to forgetfulness or physical issues. The novel architecture automatically recognizes the preparation of food at home, in a privacy-preserving and unobtrusive way, exploiting air quality data acquired from a commercial sensor, statistical features extraction, and a deep neural network. A robotic system prototype is used to simplify the interaction with the inhabitant. For this work, a large dataset of annotated sensor data acquired over a period of 8 months from different individuals in different homes was collected. Overall, the results achieved in the thesis are promising, and pave the way for several real-world implementations and future research directions.