A Stacked-Ensemble Approach to Predicting Cosmic Radiation Exposure in Commercial Aircraft using ARMAS Dataset

Key points:

1) Radiations emanating from cosmic rays are potentially high at aviation altitudes.

2) Frequent flyers and flight crew face heightened cancer risks due to elevated annual radiation exposure at aviation altitudes.

3) A machine learning-based, stacked ensemble approach is used to predict and classify the measured absorbed dose rate at aviation altitudes.

4) The model offers class probabilities, uncertainty estimation, a reliability curve, and predicts potentially dangerous radiation exposure.

Ionizing radiation, originating from primary protons and α particles beyond the solar system, interacts with Earth's atmosphere, producing high-energy subatomic particles and secondary ionizing radiation. The exposure of aircrews to cosmic radiation has long been acknowledged as a significant occupational health concern, prompting regulatory interventions by nations and aviation authorities following the guidelines outlined by the International Commission on Radiological Protection (ICRP). At cruising altitudes, atmospheric ionizing radiation, primarily high linear energy transfer (LET) radiation, poses a significant risk, capable of causing direct DNA damage and potential health hazards, such as cancer. Airline personnel and frequent travelers on such routes face heightened monthly and yearly radiation exposures, with potential career-limiting health implications. This research paper aims to predict radiation doses on an aircraft using the Automated Radiation Measurements for Aerospace Safety (ARMAS) dataset. We employed a stacked-ensemble classification model to predict measured absorbed dose rates, dD(Si)/dt (μGy/h), across various commercial flights in the United States. The performance was evaluated using standard metrics, including recall (R), precision (P), and F1-score. Our findings are twofold: multi-class classifications provide dose predictions, while a binary prediction approach identifies instances exceeding a predefined threshold, crucial for detecting potentially hazardous radiation levels. Additionally, we introduce an uncertainty score and a reliability curve to improve our understanding of model confidence and prevent overconfidence.