AI4Health: empowering telemedicine through explainable AI methods for personalized diagnosis, monitoring and treatment

Nowadays, advancing healthcare requires moving beyond traditional diagnostic, monitoring and therapeutic paradigms. Innovative artificial intelligence (AI)-based models could help integrate new technologies into everyday life. AI and biomedical signal analysis find great potential in the neuroengineering field, allowing the extraction of complex patterns from several signals, like brain, muscle, or motion activities. However, adopting these technologies is still barely used in uncontrolled settings due to several issues, such as complex acquisition setups, high inter- and intra-subject variability that needs long calibration time and the lack of transparency of AI-based methodologies. Among these, neural interface (NI), also known as brain computer interface (BCI), when input originates from the human brain, is a system that is able to assess both motor and cognitive interactions in neurorehabilitation scenarios. New solutions need to be developed to extend their application beyond laboratory environments or hospitals into telerehabilitation, fusing multiple heterogeneous data sources coming from wearable devices. From a methodological point of view, state-of-the-art approaches, typically based on single-channel feature extraction, do not consider brain connectivity features, which could potentially improve the outcome. This thesis aims to i) propose novel approaches for better understanding motor and cognitive interaction by fusing multimodal data from wearable devices and ii) develop algorithms integrating brain connectivity and eXplainable artificial intelligence (XAI) to improve interpretability and model outcomes. For these purposes, solutions for diagnosis, monitoring and treatment of neurological disease will be proposed. The first part of the thesis explores the most important steps typically used to construct a signal processing pipeline based on AI. Moreover, it discusses the application of AI and wearable technologies for home-based upper limb rehabilitation through a systematic literature review. The second part presents the results of two pipelines based on translation invariant and brain connectivity-based features to improve the BCI performance in binary and multiclass motor imagery tasks. Moreover, novel approaches to fuse and analyze wearable electroencephalography (EEG), electromyography (EMG) and inertial data are proposed. In detail, a deep learning model was developed to estimate wrist kinematics continuously from EMG and inertial measurement unit (IMU) data collected through ergonomic wearable devices. In addition, dry EEG, EMG and IMU were integrated to assess motor and cognitive interaction during a dual-task timed up and go (TUG) protocol involving groups of subjects affected by Multiple sclerosis (MS). Finally, the third part presents methodologies based on brain connectivity and XAI in support of medical diagnosis in the context of visual perception and Parkinson’s Disease (PD) classification. In the first study, a time-varying brain connectivity approach and graph analysis for studying hemispheric asymmetries in signal propagation dynamics following occipital transcranial magnetic stimulation (TMS) is proposed, providing potentially better insight into diagnostic and prognostic markers in the field of visual perception. The second study demonstrated how the combined impact of cognitive and brain reserves, based on functional connectivity measures extracted from resting-state functional magnetic resonance imaging (rs-fMRI), potentially enhances the PD classification and assists clinicians in its diagnosis. The findings achieved by this thesis underlie the potential of wearable devices and advanced signal processing pipelines to enhance the diagnosis, monitoring and treatment of neurological diseases. We focused our attention on portable devices with the aim of promoting a green economy and sustainability by reducing the need for patients to reach the hospital for periodic monitoring and follow-up.