Cross-Subject Startle Detection for Pilots on the Flight Deck Using Physiological Signals

In aviation, detecting startle responses in pilots is crucial for understanding their reactions in safety-critical situations. Startle events trigger sudden physiological and cognitive reactions, which affect decision-making and pose a potential risk in causing accidents. For the safety of passengers and crew, accurate startle detection models for pilots need to be developed. Due to the physiological response during startle events, one way of addressing this issue is to train models on recorded physiological signals. However, due to a pilot's demanding schedule, their time for collecting training data is limited. It is, therefore, imperative to develop startle detection models that are not only effective but also generalize between pilots as cross-subject classification models. We trained same-subject and cross-subject models to detect startle in a multimodal open-source dataset containing data of 18 users and compared their performance. Since the dataset is multimodal, we additionally investigated different fusion methods to examine their effects on the same and cross-subject training. Our findings reveal no statistically significant difference between the performance of same- and cross-subject classifier accuracy highlighting the success of our developed method concerning generalizability. Moreover, early fusion was shown to perform significantly better than late fusion in both, same- and cross-subject evaluations.