Neural Decoders for Hand Trajectory Estimation from M1 Activity

LSTM Model's Performance	Comparison of Models

Contributions

• Mohammed AbuSadeh: Initial Github repository.
  • Additional Models: LDA, PCA, XGBoost, Population Vector.
  • Report: Statistical Analysis of models.
• Erik Garcia Oyono: Revised GitHub repository and statistical analyses.
  • Models: Multivariate Linear Regression, Euler, Initial SVM+Kalman, Hidden State Kalman, LSTM.
  • Report: Overall structure. LSTM Introduction, Method, Results, and Discussion. Normality, Friedman, Bonferroni-Holm Methods, Results, and Discussion.
• Virginia Greco:
  • Models: Kalman, NB+Kalman.
  • Report: Introduction, Kalman, Discussion, Results.
• Helena Kosovac Godart:
  • Models: Kalman, NB+Kalman.
  • Report: Introduction, Kalman Methods, Discussion and future improvements.
• Anna Pahl:
  • Models: SVM+Kalman.
  • Additional Models: Recurrent Neural Network (RNN).
  • Report: Abstract, SVM, SVM+Angle-Specific Kalman, Discussion, Conclusion.

Overview

This repository contains implementations of various neural decoding models for brain-computer interface (BCI) applications. The goal is to predict hand movement trajectories from neural spike recordings using different machine learning and signal processing techniques.

Repository Structure

The repository follows a structured branching strategy:

• main: Contains only the final, validated model for predictions (LSTM).
• dev: Used for development and testing before merging into main.

Model Categories

Linear Regression

Contains Euler integration model. Stats and images added for comparison of initial model performance vs Euler integration.

• LinRegInit: Implements an initial linear regression model, later improved with Euler integration and comitted to LinearRegression.

Kalman Filter

Contains SVM Classifier + Kalman Filter. Stats and images added for comparison of KF vs HSKF vs SVM+KF implementation.

• KalmanInit: Implements the initial Kalman filter (KF).
• HiddenStateKF: Hidden State Implementation. Worse performance than initial KF.
• ClassifierKF: Contains SVM classifier with KF for improved accuracy vs initial KF.
  • NBClassifier: Implements Naïve Bayes classification with KF. Worse performance than SVM classifier.
  • SVMClassifier: Implements Support Vector Machine classification with KF. Commited to ClassifierKF and KalmanFilter since best model within this family.

Deep Learning

Contains final LSTM model. Stats and images added for comparison of different hyperparameters.

• LSTM: Implements Long Short-Term Memory network. Commited to DeepLearning, where it was improved, and main since best overall performance.

Usage

1. Clone the repository:

   git clone https://github.com/eg424/BCI.git
   cd BCI

2. Add the repository to your MATLAB path. You can do this within MATLAB:

   addpath(genpath('path_to_cloned_repo'));

3. Run the test script in MATLAB. Make sure you have the following files in your working directory:

• monkeydata_training.mat
• positionEstimatorTrained.mat
• testFunction_for_students_MTb.m

4. Run the test script. In MATLAB, execute:

RMSE = testFunction_for_students_MTb('YourTeamName');

This script will:

• Load the trained model parameters from positionEstimatorTrained.mat.
• Test the decoder using a held-out portion of the monkeydata_training.mat dataset.
• Plot predicted vs actual hand positions.
• Output the Root Mean Squared Error (RMSE), total runtime, average time per prediction, and a weighted performance rank.

Note: If positionEstimatorTrained.mat does not exist, the script will train a new model from scratch using positionEstimatorTraining.m.

Roadmap

• Developed an initial Linear Regression Model, later replaced with Euler's integration.
• Implemented a Kalman Filter. Tested and compared various methods for improved performance.
• Iteratively tested several Deep Learning approaches.
• Optimised and compared the top-three performing neural decoders.

Notes

• The neural data have been generously provided by the laboratory of Prof. Krishna Shenoy at Stanford University.