IEEE Solid-States Circuits Magazine - Fall 2023 - 48

Development of a compact, energy-efficient,
and high-density AI-driven system has the
potential to significantly impact the closed-loop
neuromodulation and BCI fields.
supports both multiclass detection
tasks for BCI applications (e.g., sixclass
finger movement and rest) and
binary classification for seizure and
tremor detection, using only 2.93 kB
of on-chip memory. The therapeutic
loop is completed by a compact
16-channel high-voltage-compliant
neurostimulator.
The NeuralTree system incorporates
cutting-edge circuit and algorithmic
techniques, such as end-to-end TDM,
feature approximations, and energyaware
learning, resulting in a new
class of energy and area efficiency.
The chip was successfully verified
using human epilepsy and PD datasets,
and in vivo in a rat epilepsy model,
using soft ECoG electrodes [38].
The 256-channel SoC has a small
active area of 3.48 mm2 in a 65-nm
TSMC process and achieves a system
energy efficiency of 0.227 µJ/class,
4.3-times better than state-of-theart
ML-embedded neural interfaces.
Indeed, the NeuralTree SoC uses significantly
less power and smaller
footprint per channel compared to the
state-of-the-art closed-loop systems,
while supporting 64-times more
sensing channels than the RNS and
Percept devices. Its modular architecture
allows for easy scalability and
customization based on the contributions
of different input channels
and signal modalities, and ensures
optimal performance tailored to the
unique characteristics of each disease
and patient.
The NeuralTree system outlined
above offers unique opportunities
for closed-loop AI-controlled disease
management in both established and
emerging neurological indications,
as well as future BCI systems. The
SoC's remarkably low-power consumption
and compact design favors
implantable systems by reducing in48
FALL
2023
vasiveness and prolonging battery
life. Furthermore, the incorporation
of intricate biomarkers and advanced
ML models enhances decoding accuracy.
Potential avenues for improvement
include the incorporation of
low-power front-end circuits capable
of cancelling large stimulation artifacts,
the integration of dynamicsaware
ML models to accommodate
neural signal fluctuations over time,
and the integration of wireless data
communication (for model training
and chip configuration) and wireless
power transfer circuits with the bidirectional
neuromodulation system.
Furthermore, exploring the potential
of brain-inspired neuromorphic
computing models, such as spiking
neural networks (SNNs) for disease
prediction [44], [45] and BCI control
[46] can open new horizons for
even more efficient, low-power, and
adaptable closed-loop interfaces in
the future [47], [48]. The remarkable
energy efficiency of neuromorphic
systems [45] is rooted in their utilization
of spike-based computational
paradigms, alongside the intrinsic
error resilience exhibited by SNNs
https://www.frontiersin.org/articles/10.3389/fnhum.2015.00003 https://www.frontiersin.org/articles/10.3389/fnhum.2015.00003
IEEE Solid-States Circuits Magazine - Fall 2023

Table of Contents for the Digital Edition of IEEE Solid-States Circuits Magazine - Fall 2023
