AI-Based Prediction of Fatalities in Flight Accidents: Insights from 75 Years of Aviation Accident Records

Abstract

Aviation accidents have reached a plateau in safety improvements since the late 1990s, emphasizing the need for advanced analytical approaches. This study utilizes a data-driven framework with Artificial Intelligence (AI) on a comprehensive dataset encompassing 75 years of global aviation accidents. This allows identification of long-term safety patterns often overlooked in studies restricted to specific regions or flight phases. The study aims to analyze long-term trends and predict future aviation accidents using Machine Learning (ML) classification models. This study involved web scraping Aviation Safety Network (ASN) database to compile the dataset, followed by Exploratory Data Analysis (EDA) to obtain insights. Support Vector Machine (SVM), Random Forest (RF), and Categorical Naive Bayes were employed for fatality prediction. EDA results show while the number of fatal accidents has declined, scheduled passenger service and En route flight phase show the highest occurrences proportionally. Furthermore, the maneuvering flight phase and military service have maximum likelihood of a fatal outcome. The predictive models achieved accuracies of approximately 79-80%. The SVM model, with the highest F1-score (79.85%), proved to be the most balanced in terms of specificity for non-fatal incidents and sensitivity for fatal ones. This result provides safety practitioners with a reliable framework for evidence-based decision making.