Skin-conductance (SC) reactivity is considered a bodily marker of better and poorer choice options, helping to guide decision-making in complex and uncertain situations. SC has been investigated extensively in the Iowa Gambling Task (IGT). The IGT was designed to examine decision-making performance under conditions of uncertainty and risk. Individual variability in behavioral IGT performance has been linked to various personality traits, including those of the Big Five. Skin-conductance responses (SCRs) also vary across individuals during decision-making. No studies have used machine learning techniques to predict personality from SC fluctuation during decision-making.
Aim: To develop an automatic method to recognize personality traits, based - in this study - on individual fluctuation in SC during decision-making in the IGT.
Call Girls in Lucknow Esha 🔝 8923113531 🔝 🎶 Independent Escort Service Lucknow
Arousal when making decisions predicts Big Five: A machine learning approach
1. Arousal when making decisions predicts Big Five:
A machine learning approach
Cátia Cepeda,1,3 Dina Rindlisbacher,1,2 Beatriz Esteves,3 Julian Schneider,1,2 Edouard Battegay,1,2 Lutz Jäncke,2 Hugo Gamboa,3 Marcus Cheetham1,2
1. Department of Internal Medicine, University Hospital Zurich, Switzerland,
2. University Research Priority Program “Dynamics of Healthy Aging”, University of Zurich, Switzerland,
3. Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, Portugal.
1. Introduction
Skin-conductance (SC) reactivity is considered a
bodily marker of better and poorer choice options,
helping to guide decision-making in complex and
uncertain situations. SC has been investigated
extensively in the Iowa Gambling Task (IGT).1 The
IGT was designed to examine decision-making
performance under conditions of uncertainty and risk.
Individual variability in behavioral IGT performance
has been linked to various personality traits, including
those of the Big Five.2 Skin-conductance responses
(SCRs) also vary across individuals during decision
making. No studies have used machine learning
techniques to predict personality from SC fluctuation
during decision-making.
Aim: To develop an automatic method to recognize
personality traits, based - in this study - on individual
fluctuation in SC during decision-making in the IGT.
2. Link to URPP Research Focus
Themes
• Therapeutic conflict arises due to disease-disease
interactions (DDIs) in multimorbidity and can
generate high decision uncertainty.
• Computer-based decision support tools for clinical
assessment and management of DDIs are in
development.
• The user’s personality can influence Human-
Computer Interaction and is likely to impact DDI-
related assessment and decision making.
Methods
• Approach to identify individual SC fluctuations.
• Machine learning to extract within-person SC data
for predictive modelling of personality.
• Predictive modelling techniques with potential for
application using other sources of data.
Contact
Marcus Cheetham,
marcus.cheetham@usz.ch
www.inneremedizin.usz.ch
References
1. Bechara, A., et al. (1997). Deciding advantageously before knowing the
advantageous strategy. Science, 275, 1293–1295
2. Denburg, N. L. et al. (2009). Poor Decision Making Among Older Adults Is
Related to Elevated Levels of Neuroticism. Annals of Behavioral Medicine : A
Publication of the Society of Behavioral Medicine, 37(2), 164–172.
3. Borkenau, P., & Ostendorf, F. (1993). NEO-Fünf-Faktoren-Inventar (NEO-FFI)
nach Costa und McCrae (S. 5-10, 27-28). Göttingen, Hogrefe.
Conclusion
The data indicate that the novel shape-sensitive EDA
model combined with the machine learning technique
may be applied to extract SC features that are
predictive of Big five personality dimensions. Future
work will apply the machine learning approach and
other data sources and tasks to model user
personality in Human Computer Interaction.
Fig.3. Shape-sensitive approach to EDA modelling for
individualized measurement of SCRs.
Table. 1. Extraction of SCL and SCR features from the EDA signal.
Fig. 2. Continuous EDA data acquisition during IGT (with choice,
anticipation and feedback phases in each trial).
3. Modelling SCRs and machine
learning
Steps (Fig.1)
Data acquisition
N = 54 healthy participants (M= 23.9; SD=4)
completed standard computerized IGT (Fig. 2) and
NEO-Five-Factor Inventory.3 EDA was acquired
continuously during IGT.
Shape-sensitive SCR model
We implemented a model of the EDA signal that is
sensitive to shape variability in SCRs to account for
individual differences in SC reactivity. The model also
enables disentangling of overlapping SCRs and
detection of low-amplitude events (Fig.3).
Feature extraction
We extracted a set of 9 features from each SCR.
Some features relate specifically to the choice,
anticipation and feedback phases of each of the 100
trials of the IGT (Table 1).
Feature selection and Model training
Ordinary Least Squares (OLS) linear regression, is
applied with the k-fold cross-validation method (5-
fold). A greedy forward step-wise selection algorithm
was used to define the best combination of SC
features that best predict the Big Five dimensions.
The cross-validation method splits the dataset into a
training set for model fitting and a test set for
evaluation of the selected model. The test set
simulates new data unused for model fitting and
selection.
Model evaluation
Root Mean Squared Error (RMSE), bias and slope are
used to evaluate the accuracy of the prediction of the
test set data.
Fig.1. Steps in modeling of SC responses and in machine learning
to identify the optimal subset of SC features that best predict the
Big Five dimensions: Neuroticism (N), Extraversion (E),
Conscientiousness (C), Openness to Experience (O),
Agreeableness (A).
Feature Meaning
SCL Tonic skin conductance level
SCR count Number of skin conductance reactions
% SCRs in Choice
phase
Percentage of SCRs that occur in
Choice phase
% SCRs Anticipation
phase
Percentage of SCRs that occur in
Anticipation phase
% SCRs in Feedback
phase
Percentage of SCRs that occur in
Feedback phase
SCR peak rate SCRs per minute
SCR amplitude Maximum SCR amplitude
SCR rise time Latency from SCR onset to amplitude
maximum
SCR half recovery
Time
Latency to reach half the maximum value
of the peak
Fig. 4. Model Evaluation. Representation of the prediction model
results for N, E, C, O, and A.
4. Results
Feature Extraction
The EDA model proved to be effective in separating
overlapped SCR events and modeling each event
(Fig. 3). The extracted original feature-set consists of
177 variables.
Feature Selection
Highly correlated features (Pearson corr. > 0.9) were
removed to reduce redundancy. This reduces the
feature set to 96 variables.
Model Training
The models for each of the Big Five rely on a
distinctive set of features. The number of optimal
features are 40 (N), 41 (E), 42 (C), 39 (O), 42(A).
Model evaluation
Evaluation shows good prediction performance
(Fig. 4). Prediction errors for Big Five are (predicted
vs. observed values): RMSE = 0.12 (N), 0.17 (E), 0.15
(C), 0.23 (O), 0.01 (A).
O
E C
A
N