A brief summary of the experiments and analysis performed to classify a prosthetic
leg dataset using supervised and unsupervised methods
Project Overview
This project involved performing multiple experiments to determine the best algorithm to classify the ambulation mode of the prosthetic leg while walked on over various terrains. To run these experiments I created Python modules and methods which handled:
- Data Wrangling
- Model Training and Evaluation
- Hyperparameter Tuning/Grid Search
- Visualization
Supervised Methods
- Decision trees with pruning
- Neural networks
- Boosting (AdaBoost)
- Support Vector Machines
- k-nearest neighbors
Unsupervised Methods
- Principle Component Analysis (PCA)
- Independent Component Analysis (ICA)
- Randomized Projections
- Linear Discriminant Analysis
- k-means
- Expectation Maximization
Dataset Description
The dataset was collected at the Shirley Ryan AbilityLab using a powered prosthetic leg on 2 different
able-bodied users.
During ambulation,
data were acquired simultaneously from 22 channels from mechanical sensors
embedded on the prosthesis.
The data were collected having the user complete
various ambulation tasks such as normal walking and going up and down
stairs and ramps resulting in a total of
2386 examples
(132 features for each example).
The labels for this data consist of 8 classes:
Level Walking (LW),
Toe Off (TO),
Ramp Descent (RD),
Stair Descent (SD),
Standing Toe Off (STTO),
Standing Heel Contact (STHC),
Mid-Swing (MSW),
and Mid-Stance (MST).
Detailed Analysis
The experimentation with the supervised and unsupervised algorithms gave me a better intuition on how I could leverage different tools when using unique datasets. To read the in-depth analyses done with this data, follow the links below. The foundation I built from these projects translated well to my work developing an intent recognition system for the powered prosthetic legs in the Center for Bionic Medicine.
