IEEE Solid-States Circuits Magazine - Fall 2023 - 45

decision making within the neural
implant itself, as implemented in a
number of seizure control systemon-chips
(SoCs) in recent years [2],
[29], [30], [31], [32], [36], [37], [38]. Leveraging
the power of AI in complex
pattern recognition, this framework
holds the potential for broader applications
beyond epilepsy and PD in the
future. It may be utilized to predict
and enhance cognitive performance,
detect and improve memory dysfunction,
classify and regulate mood
state, decode and facilitate intentional
movements in paralyzed patients
or amputees, and more.
Table 2 shows the frequently used
biomarkers and ML algorithms employed
for symptom detection and
motor decoding in various brain disorders
and BCI applications in the
literature. As illustrated in Table 2,
a wide spectrum of biomarkers with
varying complexity levels exists,
including features extracted in the
time, frequency, and phase domains.
These biomarkers underlie the disease
state in diverse neurological conditions
and across patients. This diversity
emphasizes the crucial requirement
for programmable SoC architectures
that can flexibly accommodate the optimal
feature set specific to an indication
and patient during runtime, while
supporting multiple classification
tasks with high accuracy and minimal
delay. In particular, achieving high accuracy
with on-chip AI models, while
simultaneously enabling low-power
operation and prolonged battery life,
In general, closed-loop stimulation offers
the potential for more effective stimulation
compared to its open-loop alternatives,
resulting in reduced side effects and improved
energy efficiency.
is a complex challenge. This requires
a careful selection of the ML model
and extensive co-optimization of algorithms
and hardware to meet both
accuracy and energy requirements.
Current closed-loop systems, primarily
designed for epilepsy and
PD, have limited channel count and
minimal on-device processing. This
compromises their accuracy and limits
their application in more complex
indications that require extensive
measurement of widespread brain
networks using large-scale ECoG
and/or LFP recording. One potential
solution is to enhance the decoding
accuracy by increasing the number
and distribution of electrodes, as
promised by technologies, such as
Neuralink. Yet to date, the processing
of large-scale neural signals has
predominantly been performed offimplant
[13], [23], [33], [39], necessitating
the transmission of massive
data streams via physical or wireless
links with dedicated power resources.
Development of a compact,
energy-efficient,
AI-driven system has the potential to
significantly impact the closed-loop
neuromodulation and BCI fields.
In the following, we introduce a
cutting-edge closed-loop neural
interface
SoC [37], [38] that leverages
modern ML and mixed-signal IC design
techniques to achieve exceptional
energy efficiency, channel density,
versatility, and intelligence, setting
new benchmarks in the field.
and high-density
NeuralTree: Redefining Efficiency,
Versatility, and Scalability in Neural
Interfaces
For closed-loop neuromodulation to
be effective, we need a precise measurement
of the source of pathological
neural activity to enable targeted
stimulation of specific brain regions
for symptom suppression. In the
context of epileptic seizure detection,
for instance, high spatial-resolution
ECoG recording allows for
precise localization of seizure foci
[40] and improves the performance
of ML-based seizure detectors. In
the management of movement disorders
like essential tremor and
PD, increasing the number of DBS
contacts can enhance tremor suppression
and minimize side effects
caused by off-target, nonspecific
stimulation. Similarly, high-resolution
TABLE 2. LIST OF RELEVANT BIOMARKERS AND ML MODELS WIDELY USED FOR SYMPTOM DETECTION AND MOVEMENT
DECODING IN NEUROLOGICAL AND PSYCHIATRIC DISORDERS.
DISORDERS AND
BIOMAKERS
Biomakers
EPILEPSY
Line length, spectral energy
features, phase-locking value,
phase-amplitude coupling Hjorth
parameters
MOVEMENT
DISORDERS
Spectral energy
features, Hjorth
parameters, fast and
slow high-frequency
oscillations (HFO),
HFO ratio, tremor
power, phaseamplitude
coupling
ML Models
MEMORY AND
PSYCHIATRIC
DISORDERS
Spectral energy features,
phase-amplitude
coupling, phase-locking
value, coherence
BCI
Spectral energy
features,
local motor
potential, Hjorth
parameters
Support vector machines, XGBoost, LightGBM, random forest, logistic regression, K-nearest neighbors, multilayer
perceptron, convolutional neural network, long short-term memory, spiking neural network
IEEE SOLID-STATE CIRCUITS MAGAZINE
FALL 2023
45
IEEE Solid-States Circuits Magazine - Fall 2023

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