Real-time Personalized Monitoring of Neurological Disorders on Wearable Systems

The terminology of neurological disorders encompasses a range of serious illnesses (e.g., epilepsy, Alzheimer's disease) leading to morbidity, disability, and stigma. Epilepsy alone affects over 50 million people worldwide, and these figures are rising as the population ages. The World Health Organization estimates that the proportion of the world's population aged over 60 will almost double, rising from 12% to 22% by 2050. Such an outlook puts additional pressure on already fragile healthcare systems. In this context, the recent growth of e-health and telemedicine is a counter-factor to reducing healthcare costs while democratizing services. In addition, the Internet-of-Things and, more specifically, wearable devices are also seen as disruptive technologies that pave the way for personalized, more precise treatments. However, this latter objective cannot be achieved without two essential requirements for health monitors: 1) early and accurate detection of pathological conditions and 2) real-time and long-term monitoring of the subject. These challenges lead to inherent trade-offs when designing wearable devices: high accuracy demands complex data processing, which normally reduces battery lifetime. In addition, wearable devices are resource-constrained systems with low processing capacity, limited memory, and small battery charges. Consequently, portable platforms require careful design regarding hardware, firmware, and data processing strategies. In this thesis, I first deal with the development and validation of an inconspicuous portable device for real-time electroencephalography (EEG) acquisition and processing, namely e-Glass. Then, I present two applications for e-Glass: 1) I tackle the problem of epilepsy detection on peripheral devices, and 2) I explore methodologies for cognitive workload monitoring (CWM) based on just a few EEG signals over the frontotemporal lobes. Regarding e-Glass development, I present the hardware and firmware solutions adopted in the e-Glass design to foster low-power characteristics. Furthermore, the system design description includes a lightweight data processing framework tailored to execute machine learning (ML)-based algorithms in resource-constrained devices. Concerning device validation, I present results for e-Glass' electrical characterization and an EEG acquisition assessment on a pilot study with twelve volunteers. The tested prototype exhibits input-referred noise inferior to 0.2 µVRMS , 23 nV/bit resolution, more than 18 noise-free bits, and a minimum 102 dB common-mode rejection rate at mains frequency. The EEG signals acquired with e-Glass show up to 0.93±0.06 correlation, on average, against signals recorded with the commercial device. Moreover, the acquired signals include well-known EEG patterns, displaying similar signal amplitude in time. With respect to epilepsy, I tackle the seizure detection problem with a focus on a solution for unobtrusive outpatient monitoring in real-time with wearable devices. In these conditions, the number of false alarms is one of the key indicators for the device's acceptance by patients. Thus, aiming to improve the discriminativeness of low-complex classical machine learning algorithms, I propose a new feature based on clinical practice: the approximate zero-crossing (AZC). A seizure classification assessment based solely on AZC features outperforms an equivalent method based on 56 classical features in the literature. The former detects 102 and