Baseline user calibration for cold-start model personalization in mental state estimation

Robust human state detection based on analysis of physiological signals requires model personalization since physiological reactions are individual. Personalization requires prior information, which is not available for a new, unknown person, i.e. in a cold-start. To overcome this, the current study proposes user calibration, which uses easily obtainable short baseline measurements to normalize physiological variables individually. Experiments were conducted on a cognitive load detection use case to determine effectiveness of the approach, required baseline duration, and the most suitable normalization function. In addition, the behavior of the model was analyzed with Shapley additive explanations to assess its trustworthiness. The results showed that user calibration always beat the non-personalized model, the optimal baseline duration was 3–3.5 min, and there were no differences between the different normalization functions. The model paid the greatest attention to the physiological phenomena found to be indicative of cognitive load in previous studies. The results encourage further evaluation of user calibration in different use cases for smart healthcare.