Personality assessment based on biosignals during a decision-making task
1. Student:
Beatriz Gonçalves Crisóstomo Esteves N.º 37130
November 2017
DISSERTATION FOR THE DEGREE OF
MASTER OF SCIENCE IN BIOMEDICAL ENGINEERING
Co-adviser:
Prof. Doctor Marcus Cheetham
Adviser:
Prof. Doctor Hugo Gamboa
2. Overview
2Overview
Aims of the Project
Introduction
- Concepts
Experiment
Methods
- Biosignals Processing
- Feature Selection
Results & Discussion
- Description of the Population
- Analysis of Predictive Models
- Biosignals Models
Conclusions
3. Aims of the project
3
Main goal: Find a relation between specific personality traits and changes in
biosignals recorded during a decision making task
• Process biosignals (ECG, EDA, BVP and pupillometry) during IGT performance
• Implement a novel EDA processing method
• Extract features from biosignals
• Predict personality traits through models constructed from the extracted
features from all biosignals
Research strategy:
Aims of the project
5. Concepts
5
Nervous System Network of neurons that sends electrical
signals that control body’s response
Brain regions associated with
decision making
Introduction
6. Concepts
6
Biosignals
Every biological event gives rise to physiological changes that can be measured,
analysed and processed
Biosignals
• ECG – electric activity of the heart
• EDA – changes in the skin conductance level
• BVP – blood volume that passes through the tissues
• Pupillometry – pupil diameter variation
Markers of
SNS or / and PSNS
activation
Introduction
7. Concepts
7
Personality Inner force that leads a person to act, to think
or to feel in a consistent way
Personality traits
• Present in all people but vary from person to person
• Basic dimensions of individual differences
Personality dimension
• Set of several traits
Five
Factor
Model
Extraversion
Agreeableness
ConscientiousnessNeuroticism
Openness to
Experience
Introduction
8. Concepts
8
Decision Making
Ability to choose between different courses of action based on their
consequences
• Complex process that includes multiple steps that are interrelated with each other
– Search for information, select a path, feedback of the decision
MaximizationTheory
Satisficers evaluate their options until one of them is good enough
Maximizers search until the best option is found
Introduction
9. Experiment
9
Iowa Gambling Task (IGT)
• The decks differ in terms of the amount of money that could be lost or won.
• It has 100 trials and after each one of them feedback of the choice is provided.
• The player should conclude that the decks associated with a high gain are
associated with a high loss and the decks with low gain are associated with a
low loss.
Demographic
data and
consent
IGT and
biosignals
Personality
assessment
1 2
43
Experiment
10. 10
Biosignals Processing (ECG)
• HR: statistical features
• HRV: statistical, geometrical,
frequency domain and non
linear features
Methods
11. 11
• SCL: statistical features
• SCR: number of peaks,
amplitude, rise time, half
recovery time
• Synchronization with IGT
Biosignals Processing (EDA)
Methods
12. 12
• BVP, BAV, IBI, pulse width:
statistical features;
• BVP range, first and
second derivative
Biosignals Processing (BVP)
Methods
13. 13
• Diameter, diameter variation,
area under the curve, blinking:
statistical features;
• Number of peaks
Biosignals Processing (Pupil)
Methods
14. 14
Feature Selection
• Pearson correlation coefficient – eliminate
correlated features
• OLS linear regression – compute the best
combination of features
• k-fold cross-validation method – evaluate the
precision of the prediction model
Model error = RMSD + bias + slope
Methods
15. Description of the Population
15
• Psychologists or university students
– from ETH and UZH
• German speakers
• 71 volunteers (18 male and 53
female) – 23.9 ± 4.2 years
Results & Discussion
16. Analysis of Predictive Models
16
Number of features from
the biosignals
Number of features after
the feature selection with
the Pearson correlation
Biosignal Subjects Testing Set
ECG 49 4-5
EDA 54 5-6
BVP 32 3-4
Pupil 67 6-7
All 24 2-3
Results & Discussion
17. ECG Models
17
O – 0.01 A – 0.02 Max – 0.02
• All scales need less than 44 features
to minimize the model error
• Best prediction error – O, A and Max
• Most chosen features are related to
the low frequencies of the HRV
• Features from Block 1 are the most
used in all personality scales
Results & Discussion
18. EDA Models
18
A – 0.05 Max – 0.05
• All scales need less than 47 features
to minimize the model error
• Best prediction error – A and Max
• Synchronization with IGT features
chosen across all personality scales
• Features from Block 1 are the most
used in all personality scales
Results & Discussion
19. BVP Models
19
O – 0.05 E – 0.03
• All scales need less than 29 features
to minimize the model error
• Best prediction error – O and E
• BAV, BVP and pulse width statistical
features chosen across all personality
scales
• Features from the complete task are
the most used in all personality scales
Results & Discussion
20. Pupillometry Models
20
A – 0.3 Max – 0.3
• All scales need less than 28 features
to minimize the model error
• Best prediction error – A and Max
• Implement features that correlate
events on the pupil diameter signal
and the moments after a loss in IGT
Results & Discussion
21. All Biosignals Models
21
• All scales need less than 21 features
to minimize the model error
• Most features are selected from the
ECG and EDA signals – most chosen
for the C, E, A and R
• Features from Block 1 are the most
used in all personality scales
Results & Discussion
22. General Discussion
22
• ECG models present slightly better results using less features than
EDA based models
• The BVP models for E and O scales show prediction errors similar to
ECG and EDA results
• O, A and Max scales have the lowest model error using features from
ECG, the C, N and R using features from EDA and the E using features
from BVP
• The O, A and R best predictive models using only features from one
biosignal present lower model errors than models with all biosignals
Results & Discussion
23. Conclusions
23
• All processing tools proved to be effective for the extraction of features
• The efficiency of the EDA model to detect low amplitude events and overlapping events was
comproved
• Predictive models which use features from all biosignals perform better than the models which
use only one biosignal
• The hypothesis that personality traits is more expressed in the start of IGT was confirmed since
the highest number of features is extracted from the Block 1 of the IGT
Future Work
• All predictive models could be improved with the introduction of features that correlate events
on IGT and events on the physiological signals
• The introduction of other biosignals, such as respiration, could improve the models accuracy
• The validation of these results should also be extended to other subjects, to verify if their
accuracy remains unchanged
Conclusions
24. Student:
Beatriz Gonçalves Crisóstomo Esteves N.º 37130
November 2017
Acknowledgements:
Professor Hugo Gamboa
Professor Marcus Cheetham
Cátia Cepeda
LIBPhys - Biosignals Group
Family & Friends