Predicting fatal helicopter accidents with missing data: Do data imputation methods affect classification performance and feature explanation?

Helicopters usually operate in complex environment and their accident reports often have missing data in information tables. However, imputation techniques for tabular data are seldom applied for aviation accident data, leading to knowledge gaps in how the choice of imputations affect accident fatality prediction and interpretation. This study collects helicopter accidents from the National Transportation Safety Board (NTSB) database and characterizes the accident features with raw missingness. Seven data imputers are employed, including Mode Imputation, Multivariate Imputation by Chained Equations with Logistic Regression/Random Forest, Support Vector Classifier, CatBoost, Generative Adversarial Imputation Net (GAIN), and Tabular Prior-Data Fitted Network (TabPFN). Imputation qualities are evaluated with artificially generated missing values. The imputed datasets are then used to train five classifiers (KNN, XGBoost, CatBoost, Random Forest, and TabPFN) to predict fatal accidents. This study also proposes a framework to select the best feature explainer by considering both prediction quality and explanation stability. Results show that features including pilot’s restraint, pilot gender, and wind speed have the highest raw missing rates. TabPFN imputer achieves the best imputation accuracy. Prediction performance for fatal accident is significantly determined by the classifier rather than imputer. However, choice of imputer or classifier-imputer combination affects the features’ importance estimated by Shapley Additive Explanations (SHAP) values. Among the classifier-imputer combinations, TabPFN classifier plus a TabPFN/SVC/MICE+RF imputer yielding the best prediction, while XGBoost classifier plus a TabPFN imputer becomes the best feature explainer.