Search

EP-4740861-A2 - HIGH-DENSITY NEURAL RECORDING, VERSATILE BRAIN ACTIVITY CLASSIFICATION, AND CLOSED-LOOP NEUROMODULATION

EP4740861A2EP 4740861 A2EP4740861 A2EP 4740861A2EP-4740861-A2

Abstract

A closed-loop neuromodulation system, including an electrode array that is implantable to a brain of a subject, analog front-end device (AFD) for selectively selecting and reading a plurality of channels from electrode array, a finite impulse response (FIR) filter for selectively filtering signals from the AFD, a feature extraction engine (FEE) operatively connected to the FIR filter, configured to selectively extract features from signals provided by the FIR filter, a tree-structured hierarchical neural network classifier for detecting disease symptoms, and a multi-channel stimulator having high-voltage (HV) drivers operatively connectable to the electrode array.

Inventors

  • SHOARAN, MAHSA
  • SHIN, Uisub
  • ZHU, Bingzhao

Assignees

  • Ecole Polytechnique Fédérale de Lausanne (EPFL)

Dates

Publication Date
20260513
Application Date
20230216

Claims (15)

  1. A tree-structured hierarchical neural network classifier for detecting disease symptoms or for controlling prosthetic devices, the neural network classifier comprising: a pruned overall network structure in which power-demanding features are pruned to reduce the number of features per node, from the overall number of features; and a plurality of internal nodes, each node represented by a 2-layer sparsely connected neural network (NN), wherein the network structure has been regularized by a power-dependent regularization during training, and wherein a single multiply-and-accumulate (MAC) and a comparator are reused for successive node processing during inference.
  2. The tree-structured hierarchical neural network classifier according to claim 1 configured to process a limited number of features on a window-by-window basis.
  3. The tree-structured hierarchical neural network classifier according to the preceding claim wherein the limited number of features are a number 64 or fewer.
  4. The tree-structured hierarchical neural network classifier according to claim 2 or 3, wherein the tree is pruned such that the maximum number of features extracted per node is limited to the limited number.
  5. A filter and feature extraction engine device operatively connected to the tree-structured hierarchical neural network classifier of any preceding claim, comprising: a time-division multiplexed (TDM) finite impulse response (FIR) filter including a bandpass filter, a Hilbert transformer, and a bypass path to selectively provide bandpass filtered signals, Hilbert transformed signals, and bypassed signals; and a time-division multiplexed (TDM) feature extraction engine (FEE) operatively connected to the FIR filter, configured to selectively extract phase synchrony features from the Hilbert transformed signals, frequency features from the bandpass filtered signals, and temporal features from the bypassed signals.
  6. The filter and feature extraction engine device operatively connected to the tree-structured hierarchical neural network classifier, according to the preceding claim, wherein the feature extraction engine (FEE) is configured to extract the phase synchrony features, the frequency features, and the temporal features one at a time.
  7. The filter and feature extraction engine device operatively connected to the tree-structured hierarchical neural network classifier, according to either of the two directly preceding claims, further comprising an accumulator configured to be shared in computing ∑ t = 1 N x t or ∑ t = 1 N x t of common mathematical expressions in different feature algorithms, among which a. Spectral Energy (SE), Local Motor Potential (LMP), Hjorth activity (ACT), Hjorth mobility (MOB), Hjorth complexity (COM), and High-Frequency Oscillation Ratio (HFOR), b. a differentiator and a further accumulator configured to be shared in computing ∑ t = 1 N x t − x t − 1 or ∑ t = 1 N Δ x t of further common mathematical expressions, among which Line Length (LL), Hjorth mobility (MOB), and Hjorth complexity (COM), and c. a ratio calculator configured to be shared in computing feature in fractional form of even further common mathematical expressions, among which High-Frequency Oscillation Ratio (HFOR), Hjorth mobility (MOB), and Hjorth complexity (COM).
  8. A closed-loop neuromodulation system, comprising: an electrode array that is implantable to a brain of a subject; an analog front-end device (AFD) for selectively selecting and reading a plurality of channels from electrode array; a finite impulse response (FIR) filter for selectively filtering signals from the AFD; a feature extraction engine (FEE) operatively connected to the FIR filter, configured to selectively extract features from signals provided by the FIR filter; a tree-structured hierarchical neural network classifier for detecting disease symptoms or for controlling prosthetic devices according to any preceding claim 1-4 ; and a multi-channel stimulator having high-voltage (HV) drivers operatively connectable to the electrode array.
  9. The closed-loop neuromodulation system of the preceding claim, wherein the AFD includes a switch matrix and dual multiplexing choppers for dynamically selecting the plurality of channels from the electrode array, and a plurality of coarse and a fine DC servo loops (DSL) configured to permit dynamic channel selection by cancelling electrode DC offsets (EDOs) that vary between successive channels, to provide EDO adjusted signals.
  10. The closed-loop neuromodulation system of the preceding claim, wherein the coarse DSL is configured to search for binary bit representations of EDOs from a group of channels, and stores them into a local memory.
  11. The closed-loop neuromodulation system of the preceding claim, wherein the fine DSL is configured to add the stored EDOs and the output of a digital integrator, delta-sigma modulate the added signals, and feeding them back to the input of an amplifier through a digital-to analog converter to remove residual EDOs.
  12. The closed-loop neuromodulation system of any preceding claim 8-11, wherein the filter and feature extraction engine device comprises: a TDM finite impulse response (FIR) filter including a bandpass filter, a Hilbert transformer, and a bypass path to selectively provide bandpass filtered signals, Hilbert transformed signals, and bypassed signals; and a TDM feature extraction engine (FEE) operatively connected to the FIR filter, configured to selectively extract phase synchrony features from the Hilbert transformed signals, frequency features from the bandpass filtered signals, and temporal features from the bypassed signals.
  13. The closed-loop neuromodulation system of any preceding claim 8-12, wherein the filter and feature extraction engine device is configured to extract the phase synchrony features, the frequency features, and the temporal features one at a time.
  14. The closed-loop neuromodulation system of any preceding claim 8-13, wherein the filter and feature extraction engine device, further comprises an accumulator configured to be shared in computing ∑ t = 1 N x t or ∑ t = 1 N x t of common mathematical expressions in different feature algorithms, among which - Spectral Energy (SE), Local Motor Potential (LMP), Hjorth activity (ACT), Hjorth mobility (MOB), Hjorth complexity (COM), and High-Frequency Oscillation Ratio (HFOR), - a differentiator and a further accumulator configured to be shared in computing ∑ t = 1 N x t − x t − 1 or ∑ t = 1 N Δ x t of further common mathematical expressions, among which Line Length (LL), Hjorth mobility (MOB), and Hjorth complexity (COM), and - a ratio calculator configured to be shared in computing feature in fractional form of even further common mathematical expressions, among which High-Frequency Oscillation Ratio (HFOR), Hjorth mobility (MOB), and Hjorth complexity (COM).
  15. The closed-loop neuromodulation system of any preceding claim 8-14, wherein the phase synchrony features, the frequency features, and the temporal features are provided to the tree-structured hierarchical neural network classifier for said detection of disease symptoms or said controlling the prosthetic device.

Description

FIELD OF THE INVENTION The present invention is directed to the field of closed-loop neuromodulation, and systems, methods, and devices for performing closed-loop neuromodulation. BACKGROUND Closed-loop neuromodulation can alleviate disease symptoms and provide sensory feedback in various neurological disorders and injuries. Energy-efficient realization of closed-loop devices with on-site classification is critical to enhancing therapeutic efficacy. Despite recent advances, existing devices, for example system-on-chip devices (SoCs), with integrated machine learning are constrained by low channel count, for example a channel count around 8-32, and poor generalizability. In light of these deficiencies of the state of the art, strongly improved methods, systems and devices for closed-loop neurostimulation are strongly desired, to provide for high channel count, low power consumption, reduced surface area usage for chip implementation, and for providing superior performance and a multitude of applications fields over the state of the art. SUMMARY According to a first aspect, the invention provides an analog front-end device for selectively selecting and reading a plurality of channels from an electrode array, the electrode array being implantable to a brain of a subject, the analog front-end device comprising: a switch matrix and dual multiplexing choppers for dynamically selecting the plurality of channels from the electrode array; and a plurality of coarse and a fine DC servo loops (DSL) configured to perform dynamic channel selection by cancelling electrode DC offsets (EDOs) that vary between successive channels, to provide EDO adjusted signals. In a preferred embodiment, the coarse DSL is configured to search for binary bit representations of EDOs from a group of channels, and stores them into a local memory. In a further preferred embodiment, the fine DSL is configured to add the stored EDOs and the output of a digital integrator, delta-sigma modulate the added signals, and feeding them back to the input of an amplifier through a digital-to analog converter to remove residual EDOs. In a further aspect, the invention provides a filter and feature extraction engine device for use with a front-end device as described herein above, comprising: a time-division multiplexed (TDM) finite impulse response (FIR) filter including a bandpass filter, a Hilbert transformer, and a bypass path to selectively provide bandpass filtered signals, Hilbert transformed signals, and bypassed signals; and a time-division multiplexed (TDM) feature extraction engine (FEE) operatively connected to the FIR filter, configured to selectively extract phase synchrony features from the Hilbert transformed signals, frequency features from the bandpass filtered signals, and temporal features from the bypassed signals. In a further preferred embodiment, the feature extraction engine (FEE) is configured to extract the phase synchrony features, the frequency features, and the temporal features one at a time. In a further preferred embodiment, the feature extraction engine device further comprises an accumulator configured to be shared in computing ∑t=1Nxt or ∑t=1Nxt of common mathematical expressions in different feature algorithms, among which Spectral Energy (SE), Local Motor Potential (LMP), Hjorth activity (ACT), Hjorth mobility (MOB), Hjorth complexity (COM), and High-Frequency Oscillation Ratio (HFOR),a differentiator and a further accumulator configured to be shared in computing ∑t=1Nxt−xt−1 or ∑t=1NΔxt of further common mathematical expressions, among which Line Length (LL), Hjortz mobility (MOB), and Hjorth complexity (COM), anda ratio calculator configured to be shared in computing feature in fractional form of even further common mathematical expressions, among which High-Frequency Oscillation Ratio (HFOR), Hjorth mobility (MOB), and Hjorth complexity (COM). In a further preferred embodiment, the phase synchrony features, the frequency features, and the temporal features are provided to a NeuralTree classifier for detection of disease symptoms. In a further aspect, the invention provides a single tree-structured hierarchical neural network classifier operatively connected to the FEE of the filter and feature extraction engine device described herein above. In a further preferred embodiment, the single tree-structured hierarchical neural network is configured to process a limited number of features on a window-by-window basis. In a further preferred embodiment, the limited number of features are a number 64 or fewer. In a further preferred embodiment, the tree is pruned such that the maximum number of features extracted per node is limited to the limited number. In a further aspect, the invention provides a closed-loop neuromodulation system, comprising: an electrode array that is implantable to a brain of a subject; analog front-end device (AFD) for selectively selecting and reading a plurality of channels from electrode array; a fi