EEG-Based Drowsiness Analysis to Establish Driving Safety Using Deep Learning Approaches

Abstract Drowsiness is the inactivated state of the brain and observed during the transition from awaken state to sleepy state. This inactive state diminishes an individual's attention and leads to accidents during professional or personal activities. The prediction of this inactive (drowsiness) state using AI plays a substantial role in the avoidance of accidents. The advancements in the field of Artificial Intelligence and Neuroscience approaches are used for the prediction of this inactive drowsy state. In order to prevent these devastating accidents, the state of drowsiness of the driver has to be be monitored. Electroencephalogram (EEG) is a predominant tool adopted to examine various states of the brain effectually. It is generally known as Brain-Computer Interface System. The EEG channels are used for predicting the inactive state while implementing the real-time applications. However, the researchers face various challenges during execution based on the classification and channel selection process. This research concentrates on modelling and efficient drowsiness prediction methods and intends to bridge the gap encountered in the existing approaches. A novel stacked Long-Term Short Memory \((s-LSTM)\) with Deep Fully Connected- Convolutional Neural Network (\(DFC-CNN\)) is used to learn and memorize the long-term feature dependencies and attains essential information based on time-series prediction. Single and multi-channel EEG data is considered to measure the statistical characteristics of available EEG signals. The online available OpenBCI sleep analysis data is used for performing the experimentation, and the simulation is done in a Pycharm and run in GoogleColab environment. The proposed \(s-LSTM\) model provides a better trade-off compared to existing approaches. The model generalization is improved with the validation of combined feature subjects. Here, metrics like prediction accuracy, RMSE, false positives, scaling coefficients related to false positives are measured to show the significance of the model.