nwang_aerospace13_machine_learning_framework.ppt

Machine learning

1. A Machine Learning Framework for Space Medicine
Predictive Diagnostics with Physiological Signals
Ning Wang, Michael R. Lyu
Dept. of Computer Science & Engineering, Chinese University of Hong Kong, Hong Kong
Chenguang Yang
School of Computing and Mathematics, Plymouth University, United Kingdom

2. Outline
Introduction
Electroencephalogram (EEG) in Aerospace Medicine
Amplitude and Frequency Properties in EEG
Predictive Diagnostics Framework
Case study: Epileptic Seizure Prediction with EEG
Discussion & Conclusion
2

3. Prognostics and health management (PHM)
3
For space missions
Focuses on fundamental issues of system failures
To predict when failures may occur
For healthcare in space
Preventive, occupational
To predict and prevent health problems timely
Subjects are pilots, astronauts, or persons involved in spaceflight
Critical to aviation safety

4. Aerospace medicine
4
Predictive diagnostics
Autonomously predict, prevent and manage potential health problems
Identify negative health trends with concerned premonitory symptoms.
Predict future health condition.
Raise alarms in case of emergency.
Disease prediction & health monitoring
Computer-based, self-diagnosis, and self-directed treatment programs
Forecast acute disease onset.
Monitor health condition.
Patient-specific.

5. EEG in aerospace medicine
5
Long been employed in crew selection andtraining.
Considered as an essential health metric of people involved in
space missions.
Diagnosis for neurologic events.
Help in determining an acute cardiovascular disease, etc.

6. How to acquire EEG data?
Data recording
Noninvasive electrodes uniformly arrayed on the scalp.
Channel signal = difference between potentials measured at two
electrodes.
Annotated to be clinical events or not by medical experts.
Scalp EEG
6

7. EEG signal’s rhythmic pattern
7

8. Amplitude and frequency properties in EEG
An EEG signal is typically described in terms of rhythmic
activities.
Contains multiple frequency components.
Differs in structure among subjects.
A band-limited signal that describes the kthEEG rhythm
is characterized by two sequences:
 --amplitudeof rhythm;
 --phaseof rhythm.
8
Extract dominant amplitude and phase components as signal descriptors,
i.e., physiological cues!

9. Observations
9
Inclusive EEG
rhythms
Estimated
frequency
components

10. Predictive Diagnostics Framework
10
Physiological signal analysis algorithm
Identify primary components
Disease prediction and health monitoring architecture
Machine learning based, subject-specific

11. Machine learning
11
“… a computer program that can learn from experiencewith
respect to some class of tasksand performancemeasure …”
(Mitchell, 1997)
“Machine learning, a branch of artificial intelligence, is about
the construction and study of systems that can learnfrom data.
For example, a machine learning system could be trained on
email messages to learn to distinguish between spam and
non-spam messages. After learning, it can then be used to
classify new email messages into spam and non-spam
folders. …”
(from Wikipedia)

12. About support vector machine (SVM)
12
Linear discriminant function
Maximal margin best hyperplane.
Support vectors: data points closest to the hyperplane.

13. Case study: epileptic seizure prediction
13
Epilepsy diagnosis
EEG with epileptic seizure
Prediction system specification
Performance

14. Epilepsy
14
Neurological disorder characterized by
sudden recurringseizures.
Affecting 1% of world’s population.
Second only to stroke.
Frequently encountered in-flight
medical events
Unpredictable time and occasions.
Second only to dizziness.

15. What happens today?
15
Diagnosis using electroencephalogram (EEG)
Recording electrical activity of brain using multiple electrodes
Machine learning techniques applied to classify EEG data
Restricted to clinical environment

16. EEG with epileptic seizure
16
Preictal–the period before seizure onset occurs.
Ictal–the period during which seizure takes place.
Postictal–the period after the seizure ends.
Interictal–the time between seizures.

17. Seizure diagnosis tasks
17
Task Requirements Application scenarios
Seizure event
detection
greatest possible accuracy,
not necessarily shortest delay.
Apps. requiring an accurate account of
seizure activity over a period of time.
Seizure onset
detection
shortest possible delay,
not necessarily highest accuracy.
Apps. requiring a rapid response to a seizure.
e.g.,initiating functional neuro-imaging studies to
localize cerebral origin of a seizure.
Seizure
prediction
highest possible sensitivity,
lowest possible false alarms,
actionable warning time.
Apps. requiring quickreaction to a seizure by
delivering therapy or notifying a caregiver,
before seizure onset.

18. Current approaches
Pattern recognition issue
Two-step processing strategy
Feature extraction front-end
Usually computationally expensive.
Standard machine learning techniques
Artificial neural networks;
Decision trees;
Mixture Gaussian models;
Support vector machine (SVM).
18
Efficient signal analysis method that can produce physically meaningful
and effective features is highly desirable!

19. Freiburg EEG database
19
Epilepsy Center, the University Hospital of Freiburg, Germany.
IntracranialEEG data:
recorded during invasivepresurgical epilepsy monitoring.
21 patients:
8 males, 13 females.
For each patient:
at least 100 min preictaldata + approximately 24 hrinterictaldata.

20. Stage Parameter Description
Data
At least 24 hrDuration of interictalrecord
At least 150 minDuration of preictalrecord
Feature extraction
5 sec EEG epoch length
6 Number of EEG channels
Training 5 fold Cross validation
SVM classification
log2γ~ [-10, 10]SVM radial basis function kernel parameter
log2C~ [-10, 10]Cost parameter
20
Classification

21. Sensitivity
95.2%: 79 out of 83
seizures predicted
successfully;
Perfect results for 16
out of 19 patients.
Specificity
0.144 FAs per hour;
Two-in-a-row post-
processing: filtering
out single positive
detection.
21
Performance

22. 22
Detailed results

23. EMG with proposed framework
23
Neuromuscular abnormality detectionand muscular fatigue
prediction.
Long-duration spaceflight and absence of gravity greatly impacts
astronauts’ neural-muscular system.
Diagnosis using electromyogram (EMG).
Indicate human’s physical status.
Reflect electrical activity produced by skeletal muscles.
Amplitude is closely related to muscle force.

24. Conclusion
24
Physiological cues as physical indicators in aerospace medicine
predictive diagnostics has been investigated.
Primary amplitude and frequency components.
A new framework for improved medical operation autonomy
during space missions has been developed.
With state-of-the-art machine learning techniques.
For disease prediction and health monitoring proposes.
On a subject-by-subject basis.
Promising epileptic seizure prediction performancein case
study has been achieved.

25. 25