CLACIR: Dataset for Cognitive Load and Canine Intervention Recognition from Physiological Signals

Emotionally intelligent machines must differentiate human emotional states in the face of subtle differences and variation in states across individuals. Large datasets are needed to train models how to classify different emotional states. Here, we introduce the Cognitive Load and Canine Intervention Recognition (CLACIR) dataset, which consists of 95.8 hours of physiological responses recorded using an Empatica E4 to cognitive load and subsequent intervention from 140 participants. Physiological data consist of 3-axis accelerometry, electrodermal activity, and photoplethysmography. To validate this dataset, we employed machine learning algorithms to classify different states in the dataset with a stratified group 10-fold cross validation. We found that Linear Discriminant Analysis best distinguished between intervention classes with an accuracy of 83.5% ± 3.7%. Accelerometry measures provided the most important features for classification, with blood volume pulse measures coming in second. The CLACIR dataset performs similarly to other smaller datasets but provides a large sampling of participants with a well-defined protocol that induces a commonly investigated mental state of cognitive load and performs well with a variety of machine learning models.