Comparative analysis of machine-learning methods for prediction of pilot performance during startle events from neuropsychophysiological features of stress resilience and cognitive task scores.

Introduction: Unexpected startle events can impair pilot perception, decision-making, and performance during time-critical flight operations, making objective predictors of individual vulnerability relevant to next-generation air combat. Methods: Fifteen military pilot cadets completed a two-phase protocol in which context-free laboratory paradigms assessed neuropsychophysiological stress-resilience features and cognitive task scores, followed by a flight-simulator task involving unexpected perturbations. Pitch-tracking error was predicted using multiple regression and machine-learning regression models evaluated with leave-one-out cross-validation, with preprocessing and feature selection performed within the training folds. Results: The baseline multiple-regression model achieved a mean absolute error of 2.55° and a correlation of approximately 0.70. Under the fixed domain-balanced feature strategy, the regression neural network achieved the best predictive accuracy, with a mean absolute error of approximately 2.07° and a correlation of approximately 0.75. However, improvement over the interpretable baseline was modest, and performance depended more on feature-selection strategy than on model type. Domain-level ablation indicated that startle-related and allostatic features contributed most to prediction. Discussion: These findings suggest that, for small and heterogeneous pilot datasets, research-guided feature selection and explainable models remain highly competitive, with limited practical benefit from more complex nonlinear approaches.