Investigating Risk Factors Associated With Fatality of Trainer Aircraft Accident: Considering Data Imbalance and Training Uncertainty

This study investigates the risk factors associated with fatalities in fixed‐wing trainer aircraft accidents using data from NTSB accident reports. The study extracts 152 binary features (temporal, flight, pilot, aircraft, meteorological, and airport) and retains 34 critical features selected by chi‐square tests and random forest recursive feature elimination (RF‐RFE). To handle the severe class imbalance (9.66% fatal accidents), this study introduces seven resampling techniques under nine sampling ratios. Each sampling technique is tested on the 11 classifiers. Prediction uncertainty due to randomness of model training processes is also evaluated. Results show that KNN with random oversampling (ratio = 0.8) achieves the highest macro F1‐score (0.7561), while AdaBoost with random oversampling (ratio = 0.2) yields the best AUC (0.8776). Uncertainty analysis identifies KNN and CNN‐1D‐Focal as the most stable models. Features’ SHAP values reveal that loss of control in‐flight, takeoff phase, and airplanes with high‐power engine(s) significantly increase fatality risk, while asphalt runways, moderate altitudes, and appropriate runway lengths mitigate fatal accident likelihood.