Neural interfaces are transforming the medical field by creating a direct communication between the brain and technology. Once a distant dream, today's neural interfaces can for instance restore hearing through auditory implants, drastically reduce tremor in Parkinson's patients, provide rudimentary vision for those who have lost sight, and so on. Moreover, advanced prosthetics are beginning to return the sense of touch to patients, closing the gap between intention and perception. Looking ahead, these systems may enable seamless speech for individuals unable to communicate and, unlock new forms of humanmachine collaboration. To this end, next-generation neurorehabilitation systems must integrate increasingly sophisticated machine learning (ML) models to handle complex classification tasks. The promise they hold, however, relies on the collection of a large amount of input data delivered by multi-channel readout circuits. To support the readout of thousands of channels while keeping resources left for ML and stimulation, these frontends must be extremely compact and energyefficient. In addition to that, future neurorehabilitation systems also likely require operating in a closed-loop fashion to enhance the therapeutic efficacy. This poses a further challenge on the design of the readout circuitry, which also needs to cope with strong stimulation artifacts that may overwhelm and corrupt the neural data to be read out. In this context, we observe that today's existing designs either achieve efficiency and compactness, but fail in closed-loop scenarios, or they support closed-loop operation at an excessive resource cost. To address this critical gap and to reconcile these demands, this research has targeted two technical goals: 1) design an ultra-compact, energy-efficient and high-fidelity readout frontend; 2) ensure robust neural signal acquisition under stimulation artifacts. The application context around this research is the study and the alleviation of chronic poststroke symptoms (the SCATMAN project), as well as the development of more generic high-channel-count closed-loop neurorehabilitation systems. In this context, we observe that today's existing designs either achieve efficiency and compactness, but fail in closed-loop scenarios, or they support closed-loop operation at an excessive resource cost. To address this critical gap and to reconcile these demands, this research has targeted two technical goals: 1) design an ultra-compact, energy-efficient and high-fidelity readout frontend; 2) ensure robust neural signal acquisition under stimulation artifacts. The application context around this research is the study and the alleviation of chronic poststroke symptoms (the SCATMAN project), as well as the development of more generic high-channel-count closed-loop neurorehabilitation systems. Using an end-to-end design approach, the same architectural features that have been designed to boost the performance of the readout frontend are exploited to make it tolerant against stimulation artifacts, overcoming the traditional trade-off. Two prototype chips have been designed, fabricated and tested to evaluate these architectural and circuit-level solutions, culminating with the design of a neural readout frontend that features a beyond-state-of-the-art compactness of only 290 μm2 per channel with a sub-μW power consuption per channel. This design has also been proven to deliver high-fidelity neural data for ML-based processing, even in the presence of stimulation artifacts, with the smallest reported area and power overhead for artifact tolerance. In summary, this work has advanced the state of the art of neural readout frontend design, proposing a holistic solution for the development of nextgeneration neuromodulation interfaces.
Marco Francesco Carlino (Fri,) studied this question.