A 2.46-mmMiniaturized Brain-Machine Interface (MiBMI) Enabling 31-Class Brain-to-Text Decoding

Mohammad Ali Shaeri, Uisub Shin, Amitabh Yadav, Riccardo Caramellino, Gregor Rainer, Mahsa Shoaran

Abstract: Recent advancements in brain-machine interface (BMI) technology offer groundbreaking solutions for individuals with motor impairments, potentially extending to speech synthesis and handwriting assistance. However, current BMIs rely on cumbersome benchtop setups equipped with resource-intensive computing units, restricting their suitability for daily use. We introduce a miniaturized BMI (MiBMI) system capable of accurate, multiclass neural decoding and high-density sensing in a millimeter-scale silicon footprint, making it suitable for next-generation implantable BMIs. A 512-channel, 31-class neural decoder employs a novel concept of distinctive neural code (DNC) driven by a class saliency model. This facilitates the precise translation of intricate neural activity into handwritten characters using a low-complexity linear discriminant analysis (LDA) classifier. The proposed decoder achieves significant improvements in memory utilization ( ${\sim } 100{\times }$ ) and computational complexity ( ${\sim } 320{\times }$ ) compared to a conventional LDA without DNCs. Moreover, MiBMI enables area-efficient 192-channel neural recording through time-division multiplexing, demonstrating its potential for fully integrated BMIs. Fabricated in a 65-nm CMOS process, the high-channel-count BMI chipset occupies a compact area of 2.46 mm2 and consumes $883~{\mu }$ W. The proposed decoder translated human intracortical neural activity into 31 characters with 91.3% accuracy, significantly enhancing the task complexity compared to previous on-chip BMIs. Furthermore, MiBMI achieved 87% accuracy in decoding the neural responses of a rat to six classes of acoustic stimuli in an in vivo experiment.