Many neurological diseases, including Parkinson Disease (PD), affect the patient's gait.
Abstract
Many neurological diseases, including Parkinson Disease (PD), affect the patient’s gait. In this project, we checked the possibility of using person’s gait in order to classify him as a patient in PD, or as a healthy person. We took samples of both PD patients and healthy people’s gait, from a lab in “Ichilov” hospital, which specialize in this subject, and tested several ways of analyzing the given data. The goal was to build a Database that when we compare subject to it, by a given algorithm, we are able to classify his disease or diseases’ progress in a good accuracy.
Today, there are labs that traces patients’ diseases’ progress by measuring the distance and timing between his steps, but cannot determine if a checked person is heal or is going to be hill at the certain disease.
The problems
The first problem is that the labs that take the sample of a person’s gait use a net of 8 sensors around each foot.
Each sample of each one of the sensors, gives the pressure that the patient use at a specific time at a specific spot with his foot.
The sensors system:
Because of this method of measurement, there were several problems with the pre-processing the data:
- What is the best way to separate a long walk in to steps:
- Identification of exceptional steps
- Identify when the subject turned or stopped
- Identify an Error in the measurement system
- How to combine the data from separated sensors
- How many sensors we need to analyze
- How to use the data from two foots
Steps measurements and separation problems:
|
Sensors samples |
Noise in the steps samples |
|
Exception steps |
Long steps |
The second problem is, finding the best algorithm to create the classification database.
Database that will yield the best classification results on any given person.
The solutions
We developed in this project some methods to preprocess the data from the sensors, and some methods to classify a given subject.
After testing several parameters, we got the best parameters that helped us to separate the walk signal in to steps with Minimum noise and lost of information.
After doing so we could combine the data from several sensors and from the two legs.
The preprocess is done according to the next scheme:
After we have got the preprocessed database, we tested in the project several algorithms and methods.
With each algorithm and method, we tried to build the classification database, i.e. a matrix of weights vectors.
With those weights and the KNN algorithm we could classified a subject.
The whole classification process with this weight matrix is described at the next scheme:
Tools
In this project we used:
1) Samples files, as text files, which were recorded by the gait lab stuff.
2) Matlab 7.0, used to analyze and build the database.
3) Excel 2003 to analyze the algorithms results.
Conclusions
As a result of running some different algorithm with some different methods, we succeeded in reaching about 75% of classification accuracy. There is a potential of improving the accuracy, by finding another method of Pre-analysis the data. There is a need to find a method of Pre-analysis with less of important step characteristic loss.
Acknowledgment
We are grateful to our project supervisor Danni Pinkovich for his help and guidance throughout this work. We are grateful to the stuff of the lab for a gait analysis in “Ichilove” hospital, for providing us the data to work on. Many thanks to the Ollendorff Minerva Center which supported this project.