Emotion Classification Using Triple Layer CNN with ECG Signals

The growing interest in recognizing human emotions from physiological signals gathered by smart wearable devices is covered in this research study. Although these gadgets have made it feasible to discreetly and continuously record physiological signals, readings can be influenced by user activity, which makes it difficult to create reliable models for wearable sensor-based emotion identification. Though EEG provides the best signal capture to identify human emotions, it significantly interferes with daily activities and is inconvenient for daily use. Electrocardiogram (ECG) signals have been shown to have potential use in emotion identification in recent studies. In this work, a triple-layered 1-D Convolutional Neural Network (CNN) model is applied to improve emotion recognition performance using electrocardiogram (ECG) signals by utilizing hyperparameter tuning methods. We have systematically explored two well-established hyperparameter tuning methods: grid search and random search. In addition, we have harnessed Bayesian optimization, a more advanced technique acknowledged for its superior performance compared to grid and random search. Bayesian optimization has been employed in conjunction with Optuna to augment its efficacy further. Upon careful evaluation, it became evident that Bayesian optimization significantly outperforms the other two techniques. Consequently, we have implemented the model utilizing Bayesian optimization with Optuna. Our results show that the triple-layered 1-D CNN model using Bayesian optimization in conjunction with Optuna performs significantly better than other neural network architectures.