Recognizing Mental Stress in Chess Players using Vital Sign Data

Recognizing Mental Stress in Chess Players Using Vital Sign Data
Christopher Eggert, Mentor: Óscar D. Lara
Department of Computer Science and Engineering
Abstract
The identification of psychological stress can provide important feedback in
performance-critical activities. We present a system for psychological stress
detection using physiological sensors during a chess match. The sensors are
inside an unobtrusive chest strap that can be worn by the player during a
match. We investigate two approaches to stress detection: classification and
clustering. By playing games on an Android phone, the system can apply
machine learning techniques to the player’s vital sign data to give important
feedback such as which moves caused the player to become stressed during
a match.
Feature Extraction
Using 5 s, 10 s, and 15 s 50% overlap time windows, the following features are extracted:
1. Statistical
features using
from acceleration
signals
(X, time
Y, Z):windows.
Features extracted
5-second 50%
overlap
• Time- and frequency-domain features, extracted from acceleration signals.
We evaluated two algorithms: Expectation Maximization and kmeans clustering. Both algorithms generated similar clusters.
2. Structural features from physiological signals (heart rate, respiration rate, skin temp, etc.)
a. The coefficients of the polynomials that best fit the signal. The Least Squares
Algorithm is applied, based on the Sum of Squared Error:
Motivation
• Affect recognition is a burgeoning area which uses computers to
automatically detect emotions.
Clustering is important because it doesn’t rely on outside
information about when the player was stressed (unlike
classification). This allows the system to report stress even when
the player may not remember being stressed.
b. Trend:
c. Magnitude of change:
Classification
For classification, we used self-reports to determine when the
player was stressed. Classification performance was evaluated
using five algorithms: Neural Networks, J48, LogitBoost, Bayesian
Networks, and Support Vector Machines.
• Users gain objective feedback about when stress occurs and what
potentially causes it.
• Implementing stress recognition on a phone is convenient and efficient.
System Overview
Clustering
Figure 5: Classification Algorithm Evaluations
Measured attributes
• Heart rate
• Respiration rate
• Breath amplitude
• Skin temperature
• Posture
• 3D Acceleration
• ECG amplitude
Figure 1: Timer App
Conclusions
Clustering is a viable method for unsupervised detection of
psychological stress, according to the assessments.
This
technique can be used to evaluate a situation and determine
when the user was stressed, without being trained. This has
numerous applications in performance enhancement and security
situations.
Figure 2: Chess App
Also indicated is that, given user feedback, the LogitBoost
classification algorithm with a 5-second time window
outperforms the other evaluated algorithms and time windows,
with a 93.56% classification accuracy.
Vital Signs:
• Heart Rate
• Respiration Rate
Figure 3: System Overview
• Posture
• collection
Breathprotocol
Amplitude
Data
• Approval
the Institutional Review Board (IRB) at USF.
• ECGof Amplitude
• Data were collected from 8 individuals while performing all binary transitions between activities, in a
• Galvanic Skin Response
naturalistic fashion.
Figure 4: Example of Clustering
These evaluations show that a mobile phone-based system can
use both supervised and unsupervised machine learning
techniques to successfully recognize mental stress in chess
players.
UNIVERSITY OF SOUTH FLORIDA