🧠 Decoding Hand Kinematics from M1 and PMd

Decoding Hand Kinematics Using Extracellular Neural Recordings from Primary Motor and Dorsal Premotor Cortices in Macaques Performing a Sequential Reaching Task

---

📖 Overview

This repository contains a complete, report-ready analysis pipeline for studying the distinct contributions of primary motor cortex (M1) and dorsal premotor cortex (PMd) during a continuous sequential reaching task.

The workflow is implemented in a cleaned and modular notebook at Project.ipynb and includes:

1. Neural data preprocessing and smoothing.
2. Planning-versus-execution directional tuning analysis.
3. Population PSTH analysis (cue-aligned and movement-onset aligned).
4. Continuous kinematic decoding with linear and recurrent models.
5. Trajectory diagnostics and GIF export for best-trial visual summaries.

---

🚀 Key Findings

1. Temporal Dynamics: Population PSTHs indicate stronger preparatory modulation in PMd and stronger movement-linked peaks in M1.
2. Decoder Comparison: The notebook benchmarks a linear Wiener-style baseline (Ridge) against a GRU-based recurrent decoder for continuous velocity decoding.
3. Continuous Information in Both Regions: Both PMd and M1 carry decodable kinematic information, with stronger decoding typically observed in M1.
4. Acceleration Challenge: Position and velocity are generally reconstructed more reliably than acceleration, which is more sensitive to noise and smoothing assumptions.

Note: exact metric values depend on dataset/session selection and split configuration.

---

📊 Dataset

The electrophysiological and behavioral data used here was originally collected by Matthew G. Perich in the lab of Lee E. Miller at Northwestern University.

Dataset source:

• CRCNS motor-cortex dataset pmd-1
• Link: https://crcns.org/data-sets/motor-cortex/pmd-1

This notebook is currently configured to run on:

• source_data/processed/MM_S1_processed.mat

It also supports PMd+M1 sessions when M1 data is present.

---

🗂️ Repository Contents

1. Project.ipynb: Main cleaned notebook for all analyses and figures.
2. Project.pre_cleanup_backup.ipynb: Backup of the pre-cleanup notebook state.
3. source_data/processed/: Processed session .mat files.
4. exports/: Generated media outputs (including GIFs when exported).
5. processing/: MATLAB helper scripts used during preprocessing workflows.

---

🛠️ Installation and Setup

# Clone the repository
git clone https://github.com/arvindini/Introduction_to_Computational_Neuroscience.git
cd Introduction_to_Computational_Neuroscience/MDSC628/Project

# Create and activate environment (conda)
conda create -n torch-env python=3.10 -y
conda activate torch-env

# Install dependencies
pip install numpy pandas scipy scikit-learn matplotlib jupyter torch pillow

If you use micromamba, create an equivalent Python 3.10 environment and install the same packages.

---

▶️ Reproducibility

1. Launch Jupyter in this directory.
2. Open Project.ipynb.
3. Select the python3 kernel in your environment.
4. Run all cells from top to bottom.

Reproducibility details:

1. Random seeds are fixed in decoder utilities.
2. Group-aware split logic is used in decoder evaluation.
3. Notebook outputs are generated directly from code execution.

---

🧪 Modeling Summary

The decoder section focuses on two families:

1. Ridge (Wiener-style) velocity decoder on sliding neural windows.
2. GRU velocity decoder with targeted hyperparameter search.

Decoded velocity is integrated to reconstruct position. Acceleration is derived and reported as supplemental.

---

📌 Notes

1. The notebook is region-aware and gracefully handles PMd-only sessions.
2. M1 analysis branches are executed only when M1 data is available.
3. Generated visual artifacts can be exported to exports/.

---

📄 Citation and Acknowledgment

If this code or workflow contributes to your work, please cite the original dataset source and acknowledge CRCNS pmd-1 and the original data contributors.