際際滷

際際滷Share a Scribd company logo

Ntokas_Vasileios_Panagiotis_Final_Demonstration

Download as PPTX, PDF
P
Pv Vasilis

This document discusses using machine learning techniques for neural signal analysis to classify signals from cuff electrodes and provide feedback to the brain. It explores feature extraction methods, dimensionality reduction, and machine learning algorithms like ANN, SVM, and decision trees. Initial classification using MAV and STD achieved good separation of stimuli. Further classification using WEKA explored logistic regression, SMO, Naive Bayes, and IBk classifiers, with IBk achieving up to 91% accuracy on a reduced feature set found using CFS. The techniques showed promise in identifying sensory stimuli but future work is needed to implement classification in real time.

1 of 16
Download to read offline
MACHINE LEARNING FOR NEURAL SIGNAL ANALYSIS Vasilis Ntokas School of Electrical and Electronic Engineering MSc Communication and Signal Processing
THE PROBLEM Disrupted communication between brain and hand Bypassing damaged nerve via electrodes and signal processing Hand movement stabilization by providing feedback to the brain BRAIN HAND
Identify patterns and extract features of the signals provided by cuff electrodes to classify them in relation to the type of sensory stimulation Feature extraction methods to be considered: MAV, VAR, FFT, STD, MAX value. Investigate dimensionality reduction techniques, like PCA and CFS. Explore various machine learning algorithms, such as ANN, SVM, Fuzzy Logic, decision trees, logistic regression. AIMS
METHODOLOGY Data collection Filtering and Normalization Discard nonessential data Transform the dataset to an ARFF file Data Pre- processing Filter application Feature extraction Feature extraction Report findings PCA CFS Dimensionality reduction (optional) Explore various algorithms Model development 10-fold cross validation Compare the results of each algorithm Model evaluation Draft the report
SOFTWARE MATLAB JAVA - Universal Java Matrix Package - jMathPlot WEKA LyX Text Editor SOFTWARE
DATA PREPROCESSING FEATURE EXTRACTION Extracted raw signal data from MATLAB files into CSV files Retained the useful data and discarded the rest For each signal, we extracted features: MAV, VAR, STD, MAX, MEAN To enrich the dataset, we created more samples by dividing each signal into pieces of 2048 data points Applied Fast Fourier Transformation to each piece and calculated the magnitudes Created the dataset in an ARFF format
FIRST CLASSIFICATION METHOD Extracted signals: 5-8 samples per foot movement 16 electrodes recording each sample Calculated the mean, the variance and standard deviation of each sample and each electrode Plot each data point according to their cluster and see if they are separable
PINKY VS DORSAL Mean absolute value Mean absolute value Standard deviation Comparison between the different stimuli as recorded by the electrodes 1 and 14 The feature that was used is MAV Easy classification along the diagonal MAV of electrode 1 STD of electrode 1 MAVofelectrode14STDofelectrode14 Comparison between the different stimuli as recorded by the electrodes 1 and 14 The feature that was used is STD Easy classification along the diagonal
THUMB VS DORSAL Mean absolute value Mean absolute value Standard deviation Comparison between the different stimuli as recorded by the electrodes 1 and 14 The features that was used MAV Easy classification along the diagonal MAVofelectrode14STDofelectrode14 MAV of electrode 1 STD of electrode 3 Comparison between the different stimuli as recorded by the electrodes 3 and 14 The feature that was used is STD Easy classification along the diagonal
THUMB VS PINKY Mean absolute value Mean absolute value Standard deviation Comparison between the different stimuli as recorded by the electrodes 5 and 12 The features that were used is MAV Easy classification along the diagonal MAVofelectrode12STDofelectrode12 Comparison between the different stimuli as recorded by the electrodes 5 and 12 The features that were used is STD Easy classification along the diagonal STD of electrode 5 MAV of electrode 1
FEATURE SELECTION Our initial dataset consisted of 1025 features: 1024 magnitudes from the FFT plus one more indicating the electrode index that recorded the signal Next, we added MAX_value, MEAN, Abs_mean, VAR and Standard Deviation. The models performance slightly increased We also tried PCA for dimensionality reduction, keeping 80% of the datas variance Another attribute selection method was used (CFS), that reduced our dataset to only 298 features and improved the accuracy of our model
SECOND CLASSIFICATION METHOD: WEKA Simple logistic on full dataset
SECOND CLASSIFICATION METHOD: WEKA IBk on reduced dataset
RESULTS Algorithm Full dataset CFS Reduced PCA Simple logistic 84.8% 372 sec 79.5% 103 sec 79.8% 472.5 sec SMO 80.1% 739 sec 79.8% 154 sec 77% 1200 sec Na誰ve Bayes 60.8% 9.3 sec 70.6% 2.8 sec 68% 3.14 sec Bayes Net 62.4% 31 sec 71.5% 6.5 sec 64% 13.4 sec IBk 70.2% 629 sec 91.1% 204 sec 60% 405 sec J48 63.8% 75 sec 61.8% 26 sec 54% 60 sec 0.00% 20.00% 40.00% 60.00% 80.00% 100.00% Simple logistic SMO Na誰ve Bayes Bayes Net IBk J48 Full dataset CFS Reduced PCA
EVALUATION First classification method Not able to separate the thumb movement from the pinky Pairs of electrodes had to be manually selected Delivered promising results Second classification method Delivered up to 90% prediction accuracy between all muscle movements CFS feature selection boosted the models performance and efficiency Simple logistic, SMO and IBk the most powerful algorithms
CONCLUSION FUTURE WORK We explored several feature selection, dimensionality reduction and data classification methods and we discovered the most appropriate for signal processing We showed that sensory stimuli can be identified with a good standard of accuracy Implement in real time

More Related Content

Ntokas_Vasileios_Panagiotis_Final_Demonstration

  • 1. MACHINE LEARNING FOR NEURAL SIGNAL ANALYSIS Vasilis Ntokas School of Electrical and Electronic Engineering MSc Communication and Signal Processing
  • 2. THE PROBLEM Disrupted communication between brain and hand Bypassing damaged nerve via electrodes and signal processing Hand movement stabilization by providing feedback to the brain BRAIN HAND
  • 3. Identify patterns and extract features of the signals provided by cuff electrodes to classify them in relation to the type of sensory stimulation Feature extraction methods to be considered: MAV, VAR, FFT, STD, MAX value. Investigate dimensionality reduction techniques, like PCA and CFS. Explore various machine learning algorithms, such as ANN, SVM, Fuzzy Logic, decision trees, logistic regression. AIMS
  • 4. METHODOLOGY Data collection Filtering and Normalization Discard nonessential data Transform the dataset to an ARFF file Data Pre- processing Filter application Feature extraction Feature extraction Report findings PCA CFS Dimensionality reduction (optional) Explore various algorithms Model development 10-fold cross validation Compare the results of each algorithm Model evaluation Draft the report
  • 5. SOFTWARE MATLAB JAVA - Universal Java Matrix Package - jMathPlot WEKA LyX Text Editor SOFTWARE
  • 6. DATA PREPROCESSING FEATURE EXTRACTION Extracted raw signal data from MATLAB files into CSV files Retained the useful data and discarded the rest For each signal, we extracted features: MAV, VAR, STD, MAX, MEAN To enrich the dataset, we created more samples by dividing each signal into pieces of 2048 data points Applied Fast Fourier Transformation to each piece and calculated the magnitudes Created the dataset in an ARFF format
  • 7. FIRST CLASSIFICATION METHOD Extracted signals: 5-8 samples per foot movement 16 electrodes recording each sample Calculated the mean, the variance and standard deviation of each sample and each electrode Plot each data point according to their cluster and see if they are separable
  • 8. PINKY VS DORSAL Mean absolute value Mean absolute value Standard deviation Comparison between the different stimuli as recorded by the electrodes 1 and 14 The feature that was used is MAV Easy classification along the diagonal MAV of electrode 1 STD of electrode 1 MAVofelectrode14STDofelectrode14 Comparison between the different stimuli as recorded by the electrodes 1 and 14 The feature that was used is STD Easy classification along the diagonal
  • 9. THUMB VS DORSAL Mean absolute value Mean absolute value Standard deviation Comparison between the different stimuli as recorded by the electrodes 1 and 14 The features that was used MAV Easy classification along the diagonal MAVofelectrode14STDofelectrode14 MAV of electrode 1 STD of electrode 3 Comparison between the different stimuli as recorded by the electrodes 3 and 14 The feature that was used is STD Easy classification along the diagonal
  • 10. THUMB VS PINKY Mean absolute value Mean absolute value Standard deviation Comparison between the different stimuli as recorded by the electrodes 5 and 12 The features that were used is MAV Easy classification along the diagonal MAVofelectrode12STDofelectrode12 Comparison between the different stimuli as recorded by the electrodes 5 and 12 The features that were used is STD Easy classification along the diagonal STD of electrode 5 MAV of electrode 1
  • 11. FEATURE SELECTION Our initial dataset consisted of 1025 features: 1024 magnitudes from the FFT plus one more indicating the electrode index that recorded the signal Next, we added MAX_value, MEAN, Abs_mean, VAR and Standard Deviation. The models performance slightly increased We also tried PCA for dimensionality reduction, keeping 80% of the datas variance Another attribute selection method was used (CFS), that reduced our dataset to only 298 features and improved the accuracy of our model
  • 12. SECOND CLASSIFICATION METHOD: WEKA Simple logistic on full dataset
  • 13. SECOND CLASSIFICATION METHOD: WEKA IBk on reduced dataset
  • 14. RESULTS Algorithm Full dataset CFS Reduced PCA Simple logistic 84.8% 372 sec 79.5% 103 sec 79.8% 472.5 sec SMO 80.1% 739 sec 79.8% 154 sec 77% 1200 sec Na誰ve Bayes 60.8% 9.3 sec 70.6% 2.8 sec 68% 3.14 sec Bayes Net 62.4% 31 sec 71.5% 6.5 sec 64% 13.4 sec IBk 70.2% 629 sec 91.1% 204 sec 60% 405 sec J48 63.8% 75 sec 61.8% 26 sec 54% 60 sec 0.00% 20.00% 40.00% 60.00% 80.00% 100.00% Simple logistic SMO Na誰ve Bayes Bayes Net IBk J48 Full dataset CFS Reduced PCA
  • 15. EVALUATION First classification method Not able to separate the thumb movement from the pinky Pairs of electrodes had to be manually selected Delivered promising results Second classification method Delivered up to 90% prediction accuracy between all muscle movements CFS feature selection boosted the models performance and efficiency Simple logistic, SMO and IBk the most powerful algorithms
  • 16. CONCLUSION FUTURE WORK We explored several feature selection, dimensionality reduction and data classification methods and we discovered the most appropriate for signal processing We showed that sensory stimuli can be identified with a good standard of accuracy Implement in real time