EEG and Machine Learning for Precision Psychiatry

Seminar slodes

1. EEG and Machine Learning for Precision Psychiatry
Marie Zelenina
Biomedical Neuroscience Lab
Institute of Biophysics and Biomedical Engineering
[email protected]
February 20, 2019
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 1 / 36

2. Motivation
ML has been gaining more and more popular in
neuroscience/psychiatry.
1
Early precision diagnosis
2
Treatment prediction outcome
3
Reinforcement learning
fMRI is great. But so is EEG!
Cheaper, so can collect more data
Sample size is crucial for ML applications!
High temporal resolution
Easier to obtain from patients with claustrophobia, anxiety etc.
Can be more informative than fMRI or can complement fMRI
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 2 / 36

3. Kaggle EEG competition
504 teams
606 competitors
17,779 entries
$25.000 prize
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 3 / 36

4. Frame Title
Talk overview
1
Paper 1 - precise diagnosis
Subjects
Data collection
Data preprocessing
Feature selection
ML algorithms comparison
results
2
Paper 2 - treatment outcome prediction (short overview)
3
BONUS: Paper 3 - feature selection
4
Methods:
Classiers:
Linear regression, Linear Discriminant Analysis
k-Nearest Neighbours
Support vector machines
Feature selection/ dimensionality reduction:
Principal component analysis
Genetic algorithm
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 4 / 36

5. EEG+ML in psychiatry 1: Precise diagnosis
Hosseinifard, Behshad, Mohammad Hassan Moradi, and Reza Rostami.
Classifying depression patients and normal subjects using machine learning
techniques and nonlinear features from EEG signal.
Computer methods and programs in biomedicine 109.3 (2013): 339 { 345.
(Journal impact factor: 2.84, SJR: 0.786)
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 5 / 36

6. Subjects
Depression (untreated): 45
Sex: 23 females, 22 males
Age:= 33.5, SD = 10.7
Diagnosis: DSM-IV + Beck Depression Inventory
Healthy controls: 45
Sex: 25 females, 20 males
Age:= 33.7, SD = 10.2
No history of psychiatric diagnoses
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 6 / 36

7. Data collection and preprocessing
Resting-state, eyes closed, 5 mins long recording
19 active electrodes, 10-20 placement
Sampling rate: 256
Bandpass lter: 0.5 { 70 Hz
Notch lter: 50 Hz
"Artifacts were inspected visually and discarded"
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 7 / 36

8. Features
1
Band power
Alpha: 8 { 13 Hz
Beta: 13 { 30 Hz
Delt: 0.5 { 4 Hz
Theta: 4 { 8 Hz
2
Repetitiveness of the signal
Detrended uctuation analysis
3
Complexity and self-similarity of the signal
Higuchi algorithm
4
Correlation dimension
Grassberger-Procaccia algorithm
5
Instability vs. predictability of signal
Lyapunov exponent
All together - 90 features extracted
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 8 / 36

9. Feature selection
1
Principal component analysis
Method to reduce feature dimensionality.
Many features measure related properties and so they are redundant.
NOT feature selection, but dimensionality reduction. We don't choose
the best features but construct new features that summarize the
information.
Remove correlations, but not higher order dependence.
Similar to ICA, but not all components are created equal!
Eigenvector, eigenvalues...
Figure:
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 9 / 36

10. Feature selection
2Genetic algorithm(performed better)
A stochastic method for function
optimization based on the mechanics of
natural genetics and biological evolution.
At each generation, a new population is
created by the process of selecting
individuals according to their level of
tness in the problem domain, and
recombining them together using
operators borrowed from natural
genetics.
This process leads to the evolution of
populations of individuals that are better
suited to their environment than the
individuals that they were created from,
just as in natural adaptation.
Figure:
Fernando Gomez and Alberto Quesada. Genetic
algorithms for feature selection in Data Analytics.
https://www.neuraldesigner.com/blog/geneticalgorithms
for
featureselection
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 10 / 36

11. ML algorithms
1
Support Vector Machine (SVM), non-linear kernel
2
Naive Bayes
3
Linear Discriminate Analysis (LDA)
4
Logistic Regression
5
K-nearest Neighbours (KNN) classier
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 11 / 36

12. ML algorithms
LDA- makes more assumptions about underlying data. Assumes the data
is normally distributed.
Logistic regression- more exible; more robust.
Figure:
results
Pohar, M., Blas, M., Turk, S. (2004). Comparison of logistic regression and linear discriminant analysis: a simulation study.
Metodoloski zvezki, 1(1), 143.
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 12 / 36

13. ML algorithms
K-nearest neighbours
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 13 / 36

14. Results. Features - EEG bands
2*Classier Features
AlphaBeta Delta Theta
KNN 70 66.6 66.6 70
LDA 73.3 70 66.6 70
LR 73.3 70 70 70
DFA Higuchi
Correlation
dimension
Lyapunov
KNN 70 73.3 76.6 70
LDA 76.6 73.3 80 73.3
LR 76.6 76.6 83.3 73.3
2*ClassierCombined features
Non-linearPower
KNN 80 73.3
LDA 86.6 76.6
LR 90 76.6
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 14 / 36

15. Discussion
Algorithms
Best performance: LR, Closely followed by LDA
This probably means most data is close to normal distrubution
Dimensionality reduction improved performance (up to 90%)
Non-linear features
Most informative feature: Correlation dimension
It evaluates variability of human brain functioning
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 15 / 36

16. Discussion
Band features
Most informative band: alpha
Alpha band = wakeful relaxation. predominantly originate from the
occipital lobe.
Electrodes where Alpha was signicantly dierent between patients
and controls:
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 16 / 36

17. Discussion
They don't say what they measure about Bands. Rel Amp? One can
only guess
Almost no information about data preprocessing
Very badly written paper
No mention of training-testing data split rate. No mention of
validation
They don't report descriptive statistics of the data.
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 17 / 36

18. EEG+ML in psychiatry 2: Treatment outcome prediction
Khodayari-Rostamabad, Ahmad, et al.
A machine learning approach using EEG data to predict response to SSRI
treatment for major depressive disorder.
Clinical Neurophysiology 124.10 (2013): 1975 { 1985.
Over 20 available antidepressants
2 out of 3 patients do not respond to the rst antidepressants
Up to 4 antidepressant treatment trials before remission
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 18 / 36

19. Johannesen, Jason K., et al.
Machine learning identication of EEG features predicting working memory
performance in schizophrenia and healthy adults.
Neuropsychiatric electrophysiology 2.1 (2016): 3.
EEG has high temporal resolution { A lot of data!
Feature selection is usually dened a-priori { PROBLEM
SOLUTION: Use machine learning for task-specic feature selection
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 19 / 36

20. Stimuli: Sternberg Working Memory Task
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 20 / 36

21. Stimuli
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 21 / 36

22. Sample
SZ: 40
Healthy: 12
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 22 / 36

23. Sample descriptive statistics
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 23 / 36

24. Data collection
Monitor: 24 LCD (1920x1200 pixels, 75 Hz refresh rate)
Subjects 100 cm from screen
Controlled light conditions
EEG equipment: 64-channel BioSemi ActiveTwo, 10-20 placement
(+ mastoid (reference), HEOG, VEOG)
Sampling rate: 1024 Hz
Bandpass lters: 0.16-100 Hz
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 24 / 36

25. Data Preprocessing
Brain Vision Analyzer
Re-referenced to average mastoid
Broadband ltered: 170 Hz
Notch lter: 60 Hz
Segmentation:
Pre-stimulus baseline
Encoding
Retention
Retrieval
Ocular artifact correction (not specied)
Bad epoch rejection (threshold +/-75V)
Time-frequency extraction:
Theta 1: centered at 4.00 Hz, range: 3.12 - 4.88
Theta 2: centered at 6.42 Hz, range: 5.01 - 7.83
Alpha: centered at 11.26 Hz, range: 8.79 - 13.73
Beta: centered at 18.53 Hz, range: 14.46 - 22.59
Gamma: centered at 40.32 Hz, range: 31.48 - 49.16
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 25 / 36

26. ML algorithm
Support Vector Machine (SVM)
Alice Zhao, "Support Vector Machines: A Visual Explanation with Sample Python Code".
https://www.youtube.com/watch?v=N1vOgolbjSc
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 26 / 36

27. Support Vector Machines
Alice Zhao, "Support Vector Machines: A Visual Explanation with Sample Python Code".
https://www.youtube.com/watch?v=N1vOgolbjSc
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 27 / 36

28. Support Vector Machines
Alice Zhao, "Support Vector Machines: A Visual Explanation with Sample Python Code".
https://www.youtube.com/watch?v=N1vOgolbjSc
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 28 / 36

29. Support Vector Machines
Alice Zhao, "Support Vector Machines: A Visual Explanation with Sample Python Code".
https://www.youtube.com/watch?v=N1vOgolbjSc
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 29 / 36

30. Support Vector Machines
Alice Zhao, "Support Vector Machines: A Visual Explanation with Sample Python Code".
https://www.youtube.com/watch?v=N1vOgolbjSc
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 30 / 36

31. Support Vector Machines
Alice Zhao, "Support Vector Machines: A Visual Explanation with Sample Python Code".
https://www.youtube.com/watch?v=N1vOgolbjSc
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 31 / 36

32. Results
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 32 / 36

33. Results
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 33 / 36

34. Results
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 34 / 36

35. References
Johannesen, Jason K., et al. (2016)
Machine learning identication of EEG features predicting working memory
performance in schizophrenia and healthy adults
Neuropsychiatric electrophysiology2.1, 3.
Hosseinifard, Behshad, Mohammad Hassan Moradi, and Reza Rostami. (2013)
Classifying depression patients and normal subjects using machine learning
techniques and nonlinear features from EEG signal
Computer methods and programs in biomedicine109.3, 339 { 345.
Khodayari-Rostamabad, Ahmad, et al. (2013)
A machine learning approach using EEG data to predict response to SSRI
treatment for major depressive disorder
Clinical Neurophysiology124.10, 1975 { 1985.
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 35 / 36

36. The End
Marie Zelenina (IBEB) EEG and ML for precision psychiatry February 20, 2019 36 / 36