Building medical neural network-based applications on a brain-inspired computing architecture

The advancement of information technology has brought the concept of smart healthcare systems into prominence. The pAvIs project endeavors to develop innovative electronics and intelligent medical sensor systems integrated with AI algorithms. The objective is to shift from a generalized "one size fits all" approach to personalized healthcare by employing sensor-based systems that can adapt in real-time to individual patients and their operating environment. The development of such systems poses several challenges, mainly achieving real-time adaptability, dealing with power consumption and form factor. This thesis addresses these limitations by building artificial neural networks models adapted to function on a distinctive brain-inspired computing architecture, surpassing the limitations of conventional Von Neumann machines and GPUs. The adapted networks are proposed as solutions for diverse applications. Two healthcare smart sensors were investigated. Adaptive sensor arrays blanket for MRI, which classifies patient movements and estimates breathing rate. This enhances MRI imaging quality, which can be affected by patient size and motion within the scanner. And an adaptive closed-loop neuromodulation device is presented for non-invasive brain stimulation. It utilizes the classification of sleep stages to trigger audio stimulation, thereby enhancing memory consolidation. The final models proposed for each task were deployed on GrAI VIP brain-inspired chip, achieving 96.57\% accuracy in movement classification, 2.61 mean absolute error in breathing rate estimation and 83.48\% accuracy in sleep stage classification with improvement of 4.6\%, 3.75\% and 12.3\% compared to reference models respectively.