Authors:
Nele Brügge
1
;
Alexandra Korda
2
;
Stefan Borgwardt
2
;
Christina Andreou
2
;
Giorgos Giannakakis
3
;
4
;
5
and
Heinz Handels
1
;
6
Affiliations:
1
German Research Center for Artificial Intelligence, AI in Medical Image and Signal Processing, Lübeck, Germany
;
2
Translational Psychiatry, Department of Psychiatry and Psychotherapy, University of Luebeck, Lübeck, 23562, Germany
;
3
Institute of Computer Science, Foundation for Research and Technology Hellas (FORTH), Heraklion, Greece
;
4
Department of Electronic Engineering, Hellenic Mediterranean University, Chania, Greece
;
5
Institute of Agri-food and Life Sciences, University Research and Innovation Center, Hellenic Mediterranean University, Heraklion, Greece
;
6
Institute of Medical Informatics, University of Luebeck, Lübeck, Germany
Keyword(s):
Stress Detection, Multiple Instance Learning, Video Analysis, Neural Networks, Machine Learning.
Abstract:
Stress detection is a complex challenge with implications for health and well-being. It often relies on sensors recording biomarkers and biosignals, which can be uncomfortable and alter behaviour. Video-based facial feature analysis offers a noninvasive alternative. This study explores video-level stress detection using top-k Multiple Instance Learning applied to medical videos. The approach is motivated by the assumption that subjects partly show normal behaviour while performing stressful experimental tasks. Our contributions include a tailored temporal feature network and optimised data utilisation by additionally incorporating bottom-k snippets. Leave-five-subjects-out stress detection results of 95.46 % accuracy and 95.49 % F1 score demonstrate the potential of our approach, outperforming the baseline methods. Additionally, through multiple instance learning, it is possible to show which temporal video segments the network pays particular attention to.