Unsupervised stress detection algorithm and experiments with real life data

Stress is the major problem in the modern society and a reason for at least half of lost working days in European enterprises, but existing stress detectors are not sufficiently convenient for everyday use. One reason is that stress perception and stress manifestation vary a lot between individuals; hence, "one-fits-all-persons" stress detectors usually achieve notably lower accuracies than person-specific methods. The majority of existing approaches to person-specific stress recognition, however, employ fully supervised training, requiring to collect fairly large sets of labelled data from each end user. These sets should contain examples of stresses and normal conditions, and such data collection effort may be tiring for end users. Therefore this work proposes an algorithm to train person-specific stress detectors using only unlabelled data, not necessarily containing examples of stresses. The proposed method, based on Hidden Markov Models with maximum posterior marginal decision rule, was tested using real life data of 28 persons and achieved average stress detection accuracy of 75%, which is similar to the accuracies of state-of-the-art supervised algorithms for real life data.