Emergency Drug Demand Forecasting in Earthquakes with XGBoost and AFT-LSTM

Forecasting emergency drug demand (FEDD) during earthquakes is essential for optimizing logistics and improving disaster response efficiency. Traditional forecasting methods mostly rely on indirect forecasting, that is, predicting the number of casualties and then extrapolating drug demand from that. However, the number of casualties in an earthquake is influenced by numerous factors, which introduces significant computational complexity for existing methods that attempt to handle these multi-faceted variables. Furthermore, drug demand forecasts must also consider seasonal and regional variations, which existing methods fail to adequately incorporate. To address these issues, this paper proposes a novel FEDD approach that integrates XGBoost and AFT-LSTM. First, a method employs XGBoost-based multi-feature extraction techniques to identify and prioritize key factors influencing earthquake casualties. Next, an AFT-LSTM model is developed to predict casualties during earthquakes, capturing temporal dynamics and the interactions among factors. Finally, a mathematical model is established to predict total drug demand, considering the number of casualties, seasonal fluctuations, and regional characteristics. Empirical analysis demonstrates that the FEDD model performs excellently in casualty prediction, offering significant advantages over traditional BP, LSTM, and Transformer models. Moreover, the model accurately predicts total drug demand based on casualty estimates. The FEDD model provides a scientific foundation for disaster management, facilitating the efficient allocation of rescue resources during earthquake disasters and enhancing overall response efficiency.