CONTINUOUS STRESS DETECTION USING A WRIST DEVICE -IN LABORATORY AND - PowerPoint PPT Presentation

CONTINUOUS STRESS DETECTION USING A WRIST DEVICE -IN LABORATORY AND REAL LIFE- Martin Gjoreski, Hristijan Gjoreski, Mitja Luštrek, Matjaž Gams http://www.fit4work-aal.eu/index.html Workshop on Mental Health: Sensing and Intervention, UbiComp 1 13 September 2016

Motivation 26 September 2 2016

Motivation Chronical stress: • raised blood pressure • bad sleep • infections • decreased performance • slower recovery EU, work-related stress costs € 20 billion a year. 26 September 3 2016

Definition ■ Definition (Ice and James) - “Stress is considered a process by which a stimulus elicits an emotional, behavioral and/or physiological response, which is conditioned by an individual’s personal, biological and cultural context ”. 26 September 4 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 5 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 6 2016

Lab stress detector – Lab data ■ Stress inducing – math task under time and evaluation pressure (200 EUR reward for motivation) ■ 21 participants x 75 minutes of data per sensor (Heart Rate, Skin Temperature, Blood Volume Pulse, Inter-beat-interval, Electrodermal Activity, Acceleration) 26 September 7 2016

Lab stress detector – Lab data 10.95 13.33 14.05 13.81 EDA BVP ACC HR ST Anxiety Labelled Data Timing score # Participants 21 Before 10.95 Age Mean 28+-4 After Easy 13.33 After Medium 14.05 No Stress 840 minutes 13.81 Low Stress 356 minutes After Hard (End) High Stress 368 minutes 26 September 8 2016

Lab stress detector – The method SENSOR FILTERING SEGMENTATION DATA FEATURE SAMPLE FEATURE EXTRACTION SELECTION DATA LABORATORY STRESS EVALUATION MODEL LEARNING DETECTOR 26 September 9 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 10 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 11 2016

Context-based stress detector – Real-life data Real-life labeled data ■ No constraints at all # 5 Participants ■ Smartphone application Age Mean 28+-4.3 – For assessing stress levels No Stress 1216 hours at random periods of the day Low Stress 70 hours – Logger for stressful events High Stress 41 hours Stress event 26 September 12 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 13 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 14 2016

The method Laboratory Real-life Lab stress Bio data Stress data Aggregate detector predictions Activity Activity Acc. data recognizer Real-life method Context-based stress detector Real-life stress Simple contexts monitoring (e.g., date-time) 26 September 15 2016

Context-based stress detector – Experiments (2) Confusion matrices and evaluation metrics for No-Context vs. With-Context classification. Each number represents an instance/event. No-Context With Context No Stress Stress No Stress Stress No Stress 3308 1630 4932 6 Stress 34 125 47 112 F1 score 0.80 0.13 0.99 + 0.19 0.81 + 0.68 26 September 16 2016

Limitations and Future work ■ Limitations – Sample size – Age structure http://www.fit4work-aal.eu/index.html ■ Future work – More real-life data – Richer context – Personalization – Energy efficiency 26 September 17 2016

Conclusions ■ We addressed the problem of stress detection in real-life. ■ Data-preprocessing, feature extraction and feature selection methods. ■ Machine learning methods for stress detection in a constrained environments. ■ Context-based method for stress detection in an unconstrained environments. ■ The key idea is to use context information. ■ Evaluated the proposed method on a real-life data. – the presented context-based method for stress detection detects (recalls) 70% of the stress events with a precision of 95%. 26 September 18 2016