Weather-Enhanced Machine Learning for Time-Resolved Risk Stratification of Clinically Managed Hymenoptera-Related...

Insect stings, particularly those caused by Hymenoptera such as wasps and bees, are frequent triggers of severe allergic reactions and anaphylaxis, yet the ability to predict short-term risk periods based on environmental conditions has not been systematically evaluated. Meteorological factors influence both insect activity and human exposure, highlighting a relevant gap in preventive risk assessment.

This exploratory single-center study was conducted in Bochum, Germany, an urban region within the Rhine-Ruhr metropolitan area. A 17-year retrospective dataset (2005-2022) of clinically treated Hymenoptera-related sting events was analyzed to explore time-resolved, weather-informed patterns using artificial intelligence (AI)-based machine learning. The study emphasizes methodological feasibility and pattern identification rather than clinical prediction. Daily weather parameters were transformed into expert-informed indicators capturing current-season and carry-over environmental conditions. A multilayer perceptron (MLP) was trained to identify periods of increased sting occurrence, and model performance was evaluated primarily using recall to capture rare-event signals.

A total of 346 clinically significant sting events were recorded. Weather variables showed strong spatial coherence across four stations and were associated with intra-seasonal clustering of sting events rather than absolute annual incidence. Exploratory analyses suggested that earlier seasonal onset correlated with higher sting counts (Pearson R = -0.52; p = 0.037). Weekly aggregation improved model performance compared with daily prediction. The cross-validated MLP showed moderate recall (0.431) and high specificity (0.86), supporting exploratory risk stratification; however, post hoc benchmarking did not demonstrate consistent superiority over simpler baseline approaches.

This study combines a long-term clinical insect sting dataset with high-resolution meteorological data to explore time-resolved, weather-informed risk patterns using machine learning. The findings demonstrate the technical feasibility of exposure-based risk stratification in a rare-event setting. However, benchmarking showed that the MLP did not consistently outperform simpler baseline approaches for binary warning of elevated-risk periods. This proof-of-concept should therefore be interpreted as exploratory and not as a stand-alone warning system, supporting further external validation in larger, multi-center cohorts before clinical, public health, or digital health implementation can be considered.