Personalized multiclass stress and cognitive load detection

Stress and cognitive load detection have focused on a binary set-up, where stress is compared to rest, and high cognitive load to lower one. A more detailed analysis could reveal the type of stress, or the moments when a person is approaching high cognitive load, rather than reached it already. In addition, the modelling efforts have focused on finding the best classification algorithm and the biosignals to measure, and other aspects of modelling like the duration of the feature windows, personalization, and model explanations have attained less attention. This study allows stress to have different types and cognitive load to have different levels or to vary continuously. Machine learning methods are investigated to assess the effects of various modelling options in this setting, different signals and features are used to find the best input data, and the influence of the features on the results are discussed. Eye metrics are given special attention since they have been less studied. The study also examines how to personalize the models with little data. The research could make stress and cognitive load detection more precise and widespread e.g. in health domain, education, and safety critical operations.