This project aims to classify daily living activities (ADLs) performed by individuals with Parkinson’s disease using inertial measurement unit (IMU) data. Data was collected from wrist and ankle sensors during various ADLs, including activities like typing, walking, and folding sheets. The goal is to create a machine learning model that can distinguish between these activities based on sensor data.
-
Data Preprocessing:
- Cleaned and organized raw sensor data.
- Removed irrelevant columns and reset the packet counter to a time-based metric.
-
Feature Extraction:
- Applied median filtering to smooth signals.
- Extracted a set of time- and frequency-domain features, such as mean, standard deviation, skew, spectral power entropy, and signal magnitude area.
-
Model Development:
- Implemented machine learning models like Random Forest and Gradient Boosted Trees.
- Used subject-level leave-one-out cross-validation (LOSO CV) to prevent data leakage and ensure robust evaluation.
- Hyperparameter tuning and feature selection were performed to optimize model performance.
-
Evaluation:
- Models were evaluated using accuracy and F1-score metrics.
- Confusion matrices helped identify ADLs that were more challenging to classify.
- The project demonstrated the potential of IMU-based sensors in accurately classifying daily living activities in Parkinson’s patients, with some activities proving more difficult to differentiate due to similar movement patterns.