Parkinson's Gait Classification

This project investigates subject-independent binary classification of Parkinson's disease from plantar-pressure gait signals in the PhysioNet Gait in Parkinson's Disease Database.

Current status

The complete analysis pipeline has been run. It includes signal inspection, fixed-window segmentation, time-domain features, frequency-domain features, gait-cycle features, left-right asymmetry features, baseline models, nested participant-grouped tuning, and final participant-level evaluation.

The final tuned Logistic Regression model uses the combined 67-feature representation and was evaluated on 165 participants using out-of-fold predictions:

Metric	Result
Balanced accuracy	0.731
Sensitivity	0.699
Specificity	0.764
ROC-AUC	0.737
False positives	17
False negatives	28

These are research results, not clinical validation. The model misses approximately 30% of Parkinson's participants and should not be used as a standalone diagnostic tool.

Validation rule

All model evaluation is participant-grouped. Windows from the same participant are never split between training and validation. Notebook 6 additionally performs nested grouped cross-validation so feature selection and hyperparameter tuning occur inside the training folds.

Notebook workflow

Run the notebooks in this order with the project Python kernel:

01_dataset_inspection.ipynb — inspect raw recordings and metadata
02_signal_segmentation.ipynb — create fixed 10-second windows
03_feature_extraction.ipynb — create time-domain features
04_baseline_models.ipynb — evaluate baseline classifiers
05_frequency_and_gait_features.ipynb — create advanced and combined features
06_feature_selection_and_tuning.ipynb — nested grouped feature selection and tuning
07_final_evaluation.ipynb — final metrics, errors, feature stability, and ROC curve

Notebook 5 creates data/processed/segmented_windows.pkl automatically from the raw recordings when the cache is missing.

Main artifacts

data/interim/ — metadata and segmentation summaries
data/processed/gait_time_domain_features.csv — time-domain feature table
data/processed/gait_advanced_features.csv — advanced feature table
data/processed/gait_all_features.csv — combined feature table
results/nested_cv_tuning_results.csv — outer-fold tuning results
results/nested_cv_participant_predictions.csv — out-of-fold participant predictions
results/final_participant_evaluation_metrics.csv — final report metrics
results/final_participant_error_cases.csv — misclassified participants
results/final_feature_stability.csv — feature selection frequency
results/figures/ — confusion matrix, ROC, spectrum, and comparison plots

Raw data and generated processed files may be excluded from version control by .gitignore. The project currently uses notebook-based implementations; the planned src/, main.py, and requirements.txt structure has not been implemented.

poonawalat/Statistical-machine-learning-for-Parkinson-s-gait-signal-classification

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Languages