Towards continuous mental stress level estimation from physiological signals

It iswell known that chronicmental stress can cause health problems. Early stress detection can help prevent these problems. We propose and compare two approaches to estimate stress level from physiology. We have measured physiological signals in three different artificial stressful conditions involving problem solving under time pressure and memorizing exercises. Rest periods were included in the protocol to avoid crossover effects over the stress conditions. The recorded signals were: electrocardiogram (ECG), respiration, skin conductance and electromyogram (EMG) of the upper trapezius muscles. About 40 min of data were recorded from 30 healthy subjects. Subjective stress levels were measured using questionnaires. We followed a feature selection process to choose 5 physiological features to be used in the analysis. A 2-minute slidingwindowwas used to extract the features by 1-s steps. The feature values were normalized to eliminate baseline and reactivity differences among subjects. The dataset was divided five times randomly in an 80% training set and a 20% test set. The different stress estimation approaches were evaluated and compared using threemetrics. First, the classification accuracy in distinguishing between stress and rest conditionswas calculated. Second and third, the root mean square error (RMSE) and the correlationwere calculated against the subjective stress levels that the subjects indicated during the protocol. Logistic regression and linear regression were applied to obtain an estimation of the stress level. The logistic regression model provided a probability between 0 and 1of a data point belonging to a stress condition. Three concepts were tested to extend the outcome towards a continuous stress level estimation. In the first method, the probability values were interpreted directly as stress levels ranging from 0 to 1. In the second method, the relative amount of time that the measurements were classified as stress condition in the past 2 min was calculated. In the third method, the average of the probability values of the past 2minutes was calculated. Linear regression was performed against subjective stress levels measured by questionnaires. For classification we chose the optimal threshold that resulted in the highest classification accuracy to classify the estimated stress levels into the known rest and stress conditions. Results are shown in Table 1. The values in the table correspond to the average numbers over the five different training and test sets. Examples of continuous stress level estimations using methods 1 and 4 are shown in Figure 1. Method 4 (linear regression) resulted in the highest classification rate and the lowest RMSE. Method 1 showed the highest correlation with the subjective stress levels. Overall, we conclude that both linear and logistic regressions are possible candidates to provide a continuous estimation of stress level. Logistic regression has the advantage that it does not need a subjective reference like questionnaires. The approach of interpreting the probability of the logistic regression model as an estimate of the stress level has, to our knowledge, not been reported before. Our results suggest that it may provide a good estimate, but this needs to be validated in further investigations.