Detection of Pre-error States in Aircraft Pilots Through Machine Learning

This study explores the feasibility of training a machine learning model to recognize pre-error signals in the anterior cingulate cortex (ACC) using Flanker test data from the COG-BCI dataset, and subsequently employing this model to detect pre-error states in aviation pilots. To address this issue, we applied various machine learning models to the dataset, including Support Vector Machines (SVM), and Random Forests, double Convolutional Neural Network (CNN) model, and a Transformer model, renowned for handling sequential data efficiently. Pilot experiments were conducted in an Airbus A320 simulator to assess real-time cognitive activity during takeoff, involving seven pilots and six engineers. Cognitive workload (CW), heart rate (HR), and pupil diameter (PD) were measured using an EEG headset, Polar H10 heart rate monitor strap, and Gazepoint GP3 eye tracker. Results from the analysis of the FLANKER dataset using various models revealed the superiority of the transformer model, with notable reductions in false negatives and a final F1 score of 0.610. Moving beyond typical study conclusions, our objective extends to assessing model applicability in a secondary domain—evaluating the classifiers’ discriminative power during takeoff procedures for aviation pilots. Despite a slight reduction in performance, the transformer model still outperforms other models in classification with an F1 of 0.578. Although there’s room for improvement in erroneous state detection, these results indicate trends in electrical brain activity that correlate with decreased behavioral performance. The transformer’s real-time performance, with an inference speed of 0.01 s, positions it favorably for Brain-Computer Interface (BCI) applications. As we anticipate increases in classifier performance with more training data and extended polling bands, this study lays the groundwork for further research in erroneous state prediction and machine learning optimization models for BCI and real-world applications.