- 10.00 – 10.45: Trial lecture (title is to be announced)
- 12.00 – 16.00: Public defence
Ordinary opponents:
- First opponent: Professor Concetto Spampinato, Department of Electrical, Electronics and Computer Engineering, University of Catania
- Second opponent: Professor Ingelin Steinsland, Department of Mathematical Sciences, NTNU
- The chair of the committee: Associate Professor Safiqul Islam, Department of Computer Science, OsloMet
Leader of the public defence: Associate professor/Head of group Boning Feng, Department of Computer Science, OsloMet
Main supervisor: Professor Pål Halvorsen, OsloMet and SimulaMet
Co-supervisors:
- Professor Michael A. Riegler, OsloMet and SimulaMet
- Professor Hugo Lewi Hammer, OsloMet and SimulaMet
Abstract
This thesis addresses fundamental challenges in developing robust predictive models for physiological signals collected through wearable devices. Through studies in sleep recognition, affect recognition, and muscle fatigue prediction, we identified and solved three key challenges in predictive modeling.
First, we addressed individual variability in physiological responses by incorporating random effects into deep learning models to account for participant-specific variations. Second, we tackled the problem of unreliable or expensive-to-obtain labels by developing a probabilistic approach using ensembles of weak labels based on binomial distribution modeling. Third, we established a statistical validation framework to ensure reliable model performance in real-world settings by distinguishing meaningful patterns from noise.
Our contributions demonstrate practical impact across multiple domains: enabling sleep detection without expensive manual annotations, continuous muscle fatigue forecasting during climbing, and enhancing diagnostic reliability through statistical validation. This work combines statistical rigor with machine learning flexibility for robust physiological predictions.
