Comprehensive Summary
This article presents a highly integrated hardware architecture designed for brain-machine interface (BMI) applications. The design implements a 1-dimensional convolutional neural network (1D CNN) feature extractor across 192 parallel channels in a 65 nm CMOS process, achieving an ultra-low power consumption of ~1.8 µW and an area of ~12801 µm² per channel. It supports configurable CNN models with 2–8 feature layers and kernel lengths up to 256, while reducing caching requirements by a factor of 5 relative to conventional architectures. The system also incorporates per-channel and per-layer power switching to optimize efficiency. The authors introduce an on-chip CNN model called FENet-66 which, when applied to neural recordings from chronically implanted micro-electrode arrays in human participants with tetraplegia, outperformed traditional spiking band power (SBP) features: showing roughly 18% higher average decoding R² and 28% better performance in the fourth year of implantation. Additionally, custom 1D-CNN kernels showed a ~10% performance boost over wavelet-based features and a 38× reduction in data-stream size. The authors validated the system in real-time online closed-loop cursor control experiments in a human subject implanted for six years, demonstrating the hardware’s viability for long-term neural implant applications
Outcomes and Implications
From a medical perspective, this technology holds significant potential for advancing neuroprosthetic systems and BMI therapies for individuals with severe motor impairments, such as tetraplegia. By integrating a high-channel-count, ultra-low-power CNN feature extractor directly on chip, the system enables more efficient, stable long-term decoding of neural signals with reduced power and area overhead—critical for implantable devices. The improved decoding performance, 18% higher than conventional features and maintained over years, suggests better fidelity and reliability in translating neural activity into control signals for prosthetic limbs, communication devices, or external robotics. The fact that the system compresses neural data by 38× means less wireless bandwidth and lower telemetry requirements, which could reduce the invasiveness and power burden of implanted systems. Moreover, the demonstrated stability over a multi-year period implies the device may better handle chronic implantation challenges like tissue encapsulation, signal drift, and hardware ageing. In sum, this work points toward more durable, high-performance BMI implants that can restore functional movement, communication, or interaction for patients with paralysis, spinal cord injury, or neurodegenerative disorders—potentially improving quality of life and independence.