Physiological sensor based affective state recognition

Abstract

With rapid advancements of Medical IoT sensors in recent years, using them to recognize an individual’s affective state has become more easily attainable. If an individual’s physiological signals are recorded while they are made to experience certain feelings, the data can be used to create a model that can recognize those feelings using the sensor data. In this paper, a system is created to use data collected from physiological sensors to predict the affective state of the individual the data is extracted from. First, the sensor data was trimmed down to just the portions where the participants experience the feeling and filtered to get rid of unnecessary features and bad data. Then, the data was processed to condense the sensor readings of the entire time a user experienced a feeling into a single row that represents that time period. Finally, the data was mapped to the feeling felt. Instead of using generic colloquial terms for emotions, more abstract notions of defining emotions were used - specifically, the Valence-Arousal-Dominance space which defines emotions using these three parameters. Using that data-set, feature selection was done to find the most important features to feed to Machine Learning Models to detect the affective state of the patient in the Valence-Arousal-Dominance space. The novelty of our research comes from the features used to predict the emotions, which include statistical representations of the raw signal data and special domain features that give further insight into the signal data from EEG and ECG.