A Comprehensive Analysis of Machine Learning and Deep Learning Models for Identifying Pilots’ Mental States from Imbalanced Physiological Data

Abstract

This study focuses on identifying pilots' mental states linked to attention-related human performance-limiting states (AHPLS) using a publicly released, imbalanced physiological dataset. The research integrates electroencephalography (EEG) with non-brain signals, such as electrocardiogram (ECG), galvanic skin response (GSR), and respiration, to create a deep learning architecture that combines one-dimensional Convolutional Neural Network (1D-CNN) and Long Short-Term Memory (LSTM) models. Addressing the data imbalance challenge, the study employs resampling techniques, specifically downsampling with cosine similarity and oversampling using Synthetic Minority Over-sampling Technique (SMOTE), to produce balanced datasets for enhanced model performance. An extensive evaluation of various machine learning and deep learning models, including XGBoost, AdaBoost, Random Forest (RF), Feed-Forward Neural Network (FFNN), standalone 1D-CNN, and standalone LSTM, is conducted to determine their efficacy in detecting pilots' mental states. The results contribute to the development of efficient mental state detection systems, highlighting the XGBoost algorithm and the proposed 1D-CNN+LSTM model as the most promising solutions for improving safety and performance in aviation and other industries where monitoring mental states is essential.