Automated sleep scoring using unsupervised learning of - - PowerPoint PPT Presentation

Automated sleep scoring using unsupervised learning of meta-features

DD221X: Degree project in Computer Science May 3rd 2016 Sebastian Olsson

What is sleep scoring?

• Judgments about a sleeping individual
• Sleep stages: REM, N1, N2, N3, Awake

Hypnogram

• Sleep stage graph

Electroencephalogram (EEG)

Electroencephalogram (EEG)

Electroencephalogram (EEG)

N3

Electroencephalogram (EEG)

N3 N1

Electroencephalogram (EEG)

N3 N1 REM

Electroencephalogram (EEG)

N3 N1 REM Awake

Automated sleep stage scoring

Automated sleep stage scoring

Compare 100 % agreement

Automated sleep stage scoring

Problem statement

A

Problem statement

B

Problem statement

• Deep belief net (DBN)
• Compare approaches

B

Method

• SHHS1
• 10 records
• Annotations

Data

Segmentation

30 s

Feature extraction

(12, 0.8) (15, 0.3) Mean Variance

Features

• Mean
• Variance
• Skewness
• Kurtosis
• Hjorth mobility
• Hjorth complexity
• Amplitude

Partitioning

• 75 % training set
• 25 % test set

Feature selection

• Find a decent combination of features
• Strip away unwanted features

○ Curse of dimensionality

• Inspired by Löfhede [1]
• Genetic algorithm

○ Roulette-wheel selection ○ Mutation rate: 0.2 ○ Crossover rate: 1.0 ○ Number of generations: 5 ○ Population size: 5 ○ Chromosome length: 7

• “Cross-validation”

Feature classification

• Support vector machine (SVM)

○ Linear kernel ○ Radial basis function (RBF) kernel

• Trained using the training set
• Evaluated using the test set

Unsupervised DBN processing

• DBN

○ Two stacked Restricted Boltzmann machines (RBM)

• Based on Längkvist [2]

1. Pre-training 2. Unsupervised fine-tuning with backpropagation 3. Propagate the feature space through the network

Unsupervised DBN processing

• Three meta-features
• Appended to vector: (x1, ..., x7) ↦ (x1, …, x7, m1, m2, m3)

Evaluation

• 10 ∙ 3 ∙ 2 ∙ 2 = 120 evaluations

# records # re-runs linear/RBF kernel with/without DBN

Results

Results

Scorer A, linear kernel Scorer B, linear kernel Scorer A, RBF kernel Scorer B, RBF kernel

Results

Results

Conclusion

• Unsupervised DBN processing did not help
• Effect too small to be noticable
• Approach too specific to arrive at a general conclusion

Future work

• Simplify

○ Skip feature selection

• Append/replace
• Try different parameters, e.g.

○ Number of meta-features (output nodes) ○ Number of RBMs ○ Number of hidden layer units ○ Epochs ○ Initial biases

References

• [1] J. Löfhede. (2009). The EEG of the neonatal Brain - Classification of

Background Activity.

• [2] M. Längkvist. (2012). Sleep Stage Classification Using Unsupervised

Feature Learning.

Images

• Licensed under CC BY-SA 3.0:

○ https://commons.wikimedia.org/wiki/File:HYPNOGRAM_created_by_Natasha_k.jpg

• Licensed under CC BY-SA 4.0:

○ https://commons.wikimedia.org/wiki/File:Sleep_scoring.png

• Public domain:

○ https://commons.wikimedia.org/wiki/File:1st-eeg.png ○ https://pixabay.com/en/scientist-professor-man-researcher-28748/ ○ https://pixabay.com/en/artificial-intelligence-155161/

• Licensed under CC0 1.0:

○ All remaining images

The End