Detecting Human Emotions Through Physiological Signals Using Machine Learning

Abstract

Identifying and predicting emotions based on data can in fact sabotage many mishaps in an early stage. In this research work, we have considered four physiological signals—body temperature, heart rate, skin resistance, and pulse wave. These signals are obtained from a skin temperature sensor, a heart rate sensor, a galvanic skin response sensor (GSR), and a custom-designed pulse wave sensor. The signals are processed using a microcontroller. The microcontroller transmits the data to a computer via USB. The data is classified using machine learning algorithms for detecting various emotions. The four basic emotions considered are normal (relaxed), happy, sad, and angry. We have collected data from 22 healthy individuals, including male and female, with ages ranging from 20 to 22 years. The performance of the different machine learning algorithms on the dataset is checked through Weka and TensorFlow. The detailed results are discussed under the section Results and Discussion. Random forest tree proved to provide the highest accuracy of 82.55% for the combined dataset and 99.9% for the individual dataset, using Weka. Also, we have achieved an accuracy of 98.75% for an individual dataset through a fully connected 10 hidden layered neural network, using TensorFlow.