COMSOLE: Smart Insoles - 2021

Highlights of my works

• PD patient classification: 98.6% accuracy. 

• Disease stage classification: 97.2% accuracy.

Supervisors

• Dr. Ya Wang (Faculty at Texas A&M University)

• Rui Hua (Ph.D. candidate at Texas A&M University)

Team Members

• Songlin Xie, Malcolm Shimkus, Bryan Ham, Ha Eun Kim, Blake Smith-Cortez

Objective

Background

Parkinson's disease (PD) is a progressive nervous system disorder, and currently, there is no effective way to examine this disease. To help effectively diagnosing PD, Dr. Wang and her laboratory developed smart insoles called COMSOLE.  COMSOLE project is a long-term project. The ultimate goal is to allow individual patients to independently monitor PD progression. 

Our Deliverables

• Prototyped electronic insoles

• A signal processing model

• A user interface (UI)

Expectation

A PD patient will wear the insoles, follow the UI instructions, complete the motion experiment, and received the disease status results.

PCB

• Soldering

• Encoded acceleration data

Fabrication

• Epoxy Encasing

• Dynamic pressure analysis

User Interface

• The interface of inputting user information

• The interface of collecting experiment data

Experiment Instruction

• Toe-tapping animation

• Walking while holding a cup of water animation

• In-phase heel pivoting animation

Data

• Raw data of toe-tapping acceleration

• Training data by concatenating signals from every axis 

Machine-Learning Details

Data Collection

1. Ask volunteers to make feet gestures. 

2. Collect feet acceleration data.  

3. (16 volunteers are PD patients, and 11 volunteers are non-PD controls). 

Preprocessing

1. Decode signal data from TXT files.

2. Apply a low-pass filter to the signal series.

3. Centering the signal series by removing the mean.

4. Normalize the signal series by rescaling the signal amplitude.

5. Divide each signal series into three sample series, and concatenate one sample series from every axis.

Training

1. Use a shapelet-based classifier.

2. Use a random forest classifier.

Testing

1. Perform cross-validation in 5 folds. 

Identifying Significant Shapes

• Shapes that classify PD patients

• Shapes that classify PD stages

Classification Results

PD patients classification

•  98.55% accuracy

• Identify whether a person is a PD patient.

• Classes: Yes / No.

PD stages classification 

• 97.22% accuracy.

• Identify the PD stages of a person.

• Classes: Stage 0-4; Stage 5 is excluded.