Parkinson's Hand Tremor Prediction using ML

This is a project under NeuroTech UofT that aims to detect Parkinson's tremor using machine learning. The model will be integrated with hardware devices to intervene with tremors in PD patients. The project is divided into two parts: the first part uses traditional ML methods (using scikit-learn and xgboost) to classify tremor data, while the second part uses deep learning methods (using pytorch).

Method 1: Traditional ML

The first part of the project follows this research paper in preprocessing the public accelerometer and EEG data, then using simple ML tools such as KKN and SVM to classify it into "pre-tremor", "tremor" and "non-tremor" classes.

Data Preprocessing

First, we use the accelerometer data to extract tremor offset times, which are labels to be used in the classification because event labels are not given. This is implemented in preprocessors/acc_preprocessor.py Please refer to the paper for the complete algorithm:

The EEG data is first preprocessed by filtering and applying ICA. Then, we use the extracted timestamps for tremor onset and offset to label and segment the EEG data into epochs with "pre-tremor", "tremor" and "non-tremor" classes. The results are saved in the processed/eeg_data directory. This is implemented in preprocessors/eeg_preprocessor.py.

Feature Extraction

After preprocessing the data, you can load the eeg_data from processed/eeg_data using the EEGDataLoader class in preprocessors/eeg_loader.py. This class contains all the feature extraction methods and provide methods to obtain the features and labels for classification directly.

The following features are then extracted from each segment:

Classification

Finally, we train a classifier using the extracted features to predict the labels (pre-tremor, tremor, control). We used XGBoost in models/xgb_model.py.

Method 2: Deep Learning

Preprocessing

The second part of the project leverages latest advancements in deep learning to compare model performance with traditional ML methods. We rely on a different dataset in new_data directory. The data is first segmented into sliding windows and classified into non-tremor, pre-tremor, and tremor classes based on the given events label (more information are given in experiments/create_windows.ipynb).

No manual extraction of features are conducted in this part.

Model

We train a CNN-LSTM model to classify the segmented data. The model is implemented in models/cnn_lstm_1d.py. The model is trained using the experiments/train_cnn_lstm.ipynb notebook on Google Colab. We analysed the classification results using metrics like accuracy, precision, recall, and F1-score and training and validation loss in the results directory.

Documentation

Repository structure:

.
├── experiments/ # Jupyter notebooks for data analysis, training...
├── data/ # Original dataset files (.fif)
├── models/ # Models and helpers (e.g. loaders)
├── new_data/ # New dataset files (.bdf)
├── preprocessors/ # EEG and Accelerometer data preprocessing
├── processed/ # Processed data outputs
├── results/ # Model evaluation results
└── weights/ # Saved model weights

1. Data Preprocessing: Cleaned EEG data from original dataset are stored in processed/eeg_data directory. If it is not complete you should generate them yourself by running preprocessors/run_preprocess.py.

2. New dataset preprocessing: The new dataset should be downloaded from Google Drive instead. After that, the create_windows.ipynb notebook helps to clean and segment and save it as a pytorch tensor file.

3. Running models: Use models from models and refer to notebooks in experiments to train and evaluate the models. If you are using the new dataset, models/loader.py provides some helper utilities.

4. Results: Selected model weights are saved and uploaded to weights, as well as training results in results. The naming convention is *_old for the original dataset and *_new for the new dataset.