Parkinson-Disease

Index

1. Dataset description
2. DataClass
3. Baseline models
4. Initial Models

4.1. Motif Search

4.2. Classification

4.3. Clustering

4.4. Models on sequence data

4.5. RFClassifier

5. The Hybrid models

6. Explinable Models

---

1. Dataset Description

For the later tasks, the following data set was used. The database contains measures of gait from 93 patients with idiopathic PD, and 73 healthy controls. The dataset contains:

1. Vertical ground reaction force records of subjects as they walked for approximately 2 minutes on level ground.

• The file contains the measures from 8 sensors for each foot.

• Each individual walks for 2 minutes, records are taken at 100 samples per second.

Thus, we have 12000 record for each 2 mins walk.

2. Demographics file contains demographic information, measures of disease severity and other related measures.

2. DataClass

The dataclass contains read the data, segment it, scale it and iterpolate it.

3. Baseline models

Applying random forest classifier on the statics from the raw signal data AFTER being filtered.

4. Initial Models

4.1 Motif Search

In the first file, tried to apply motif identification on different features from the time series dataset, it appeared that the shape of the dataset (pressure - no pressure) is what resulted as motif and that is not useful.

In the second file, Tried to filter the data, but still having the same problem.

4.2 Classification

Task	features	Accuracy
Severity Detection	Univariate classification (L2 sensor)	0.39
Parkinson’s Classification	Univariate classification (L2 sensor)	0.71
Severity Detection	Multivariate classification	0.38
Parkinson’s Classification	Multivariate classification	0.82

4.3 Clustering

Clustering for fait time series dataset did not result in promising results. The first file tried to cluster the data from both lef and right feet. The second file applies the clustering on data from the left foot only. The third file applies the clustering on data from the right foot only.

4.4 Models on sequence data

Data Class + Applying FWHM on accumalated forces fro the right foot + applying models on the sequences.

4.5 RFClassifier

• RFClassifier Class for training, predicting, scoring the results with Random Forest Classifier.
• RFC models In this code file, we use class data and RFC class to apply the previous different models on the data. (SUMMING UP)

1. RFC basemodel using statics from raw data.
2. RFC models on interpolated scaled stances with FWHM algorithm, and additional features
3. RFC on scaled stances with 3 extra features and statics
4. RFC on scaled stances with 6 extra features and statics
5. RFC model on scaled stances from the right & left foot with 6 extra features each
6. RFC model on scaled stances from the right & left foot with 6 extra features each and basemodels
7. RFC model on scaled stances from 16 sensor from the right & left foot with 6 extra features each
8. RFC model on scaled stances from 16 sensor from the right & left foot with the sum of all sensors and with 6 extra features each
9. RFC model on scaled stances from 16 sensor from the right & left foot and the sum of all sensors and with 6 extra features each foot and the statics from base models

Model	Input data	Accuracy	Precision	Recall	F1
RFC n_est=200	Statics on filtered raw data from each sensor	0.8329	0.8716	0.9084	0.8844
RFC n_est=200	Right foot related: [ Interpolated Scaled stances, 3 features]	0.7092	0.7708	0.8176	0.7897
RFC n_est=200	Right foot related: [ Interpolated Scaled stances, 6 features]	0.7341	0.7896	0.8370	0.8815

5. Hybrid models

• Trying hybrid models trying different hybrid models for right stances and 3 features.

• Hybrid model Hybrid model class and a train it on for right stances and 3 features and on for right stances and 6 features. Two final notebooks for training hybrid model on all data with/without statics for binary classification and severity Detection.

6.Explinable Models

First file: Finding what features are more important on models trained on statics and extracted features

Second file: Showing the most important features in different plots.