Hierarchical Neural Interface for Continuous Motor Decoding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current direct neural interfaces face challenges in continuous decoding of motor intentions from electrocorticographic (ECoG) signals, particularly in distinguishing active and rest periods, leading to asynchronous control limitations and discontinuities in movement decoding, which restricts their application to controlling multiple limbs effectively.

Innovation Solution

A direct neural interface using a hierarchical Hidden Markov Model (H2M2) with a tree structure, combining a mixture of experts and REW-NPLS regression for calibration, allows for continuous decoding and stable control of multiple effectors by modeling the evolution of observation tensors and estimating control signals from electrophysiological inputs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If continuous decoding is implemented to control movement during any time window, then the freedom of movement control is improved, but the accuracy of distinguishing active and rest periods deteriorates

Engineering Contradiction:
Improvefreedom of movement controlVSAvoidaccuracy of distinguishing active and rest periods
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent segments the decoding process into multiple independent expert models, each specialized in decoding specific movement types or neural patterns. This segmentation allows the system to maintain high accuracy for distinguishing active and rest periods by assigning specific experts to specific decoding tasks, while still enabling continuous control through the collective operation of multiple experts.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a hierarchical dimension to the decoding architecture by organizing experts into a hierarchy with parent-child relationships. This hierarchical structure adds a new dimension to the decoding process, allowing the system to distinguish between different levels of movement complexity and maintain accurate period detection while enabling continuous control through multi-level processing.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If synchronous control is used with time windows to improve decoding accuracy, then the precision of movement control is improved, but the duration of available control time is reduced

Engineering Contradiction:
Improveprecision of movement controlVSAvoidduration of available control time
Core Design Contradiction:
Measurement precisionVSDuration of action of moving object

Solution Approach 1:

The patent enables continuous decoding by allowing multiple expert models to operate simultaneously and continuously, rather than sequentially in discrete time windows. The useful action of decoding continues uninterrupted as new neural data becomes available, with the system constantly updating movement estimates based on current neural activity patterns.

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

The patent performs preliminary decoding actions by pre-training multiple expert models during a calibration phase before actual use. These pre-trained experts are ready to immediately process new neural data without requiring re-calibration during operation, enabling continuous control while maintaining precision through the accumulated knowledge of pre-trained models.

Inventive Principle:
Principle #10Preliminary action

3Ease of operation

If a simple Markovian model is used for state transition to reduce computational complexity, then the ease of processing is improved, but the ability to model complex movement patterns deteriorates

Engineering Contradiction:
Improveease of processingVSAvoidability to model complex movement patterns
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent segments the complex movement modeling task into multiple independent expert models, each handling specific movement patterns. This segmentation maintains computational efficiency by processing each expert's predictions independently while collectively capturing complex movement behaviors through the combination of specialized models.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements a nested hierarchical structure where parent expert models oversee child expert models. This nesting arrangement allows the system to model complex movement patterns through hierarchical organization while maintaining ease of processing by allowing each level to operate semi-independently with clear parent-child relationships governing the flow of information and control.

Inventive Principle:
Principle #7Nested doll (Nesting)

4Ease of manufacture

If single-limb decoding is implemented to simplify the system architecture, then the ease of manufacture is improved, but the versatility for controlling multiple limbs is reduced

Engineering Contradiction:
Improveease of system architectureVSAvoidversatility for controlling multiple limbs
Core Design Contradiction:
Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal decoding framework where the same hierarchical expert model architecture can be applied to decode multiple limbs simultaneously. Each limb's movement can be decoded using the same expert model structure, making the system versatile for controlling multiple limbs while maintaining ease of manufacture through standardized architecture.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent segments the multi-limb control capability into independent expert models that can be instantiated for each limb. This segmentation allows the system to maintain a simple, standardized architecture for each limb while achieving versatility through the parallel operation of multiple identical or variant expert models.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4088659A1Direct neural interface using a semi-markovian hierarchised model
Publication Date: 2022.11.16 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP4088659A1 patent drawingFigure 1
  • EP4088659A1 patent drawingFigure 2A~2B
  • EP4088659A1 patent drawingFigure 3A~3B

AI summary

The present invention relates to a direct neural interface for estimating a control tensor from an observation tensor obtained by preprocessing a user's electrophysiological signals. The evolution of the observation tensor over time is modeled by an H2M2 model, comprising a plurality of HMM sub-models organized in a hierarchical tree structure. This tree structure has at its root a main HMM sub-model comprising a resting state from which different branches originate, for example, a first branch of the tree associated with a right lateral state and a second branch of the tree associated with a left lateral state.The direct neural interface uses a mixture of experts, each expert being associated with a production state of the H2M2 model, each expert (Ek) being defined by a multilinear predictive model, the control tensor being estimated by combining the predictions of the different experts. According to one variant, at least some of the submodels of the tree are semi-Markovian.