Accurate and efficient prediction of emergency material demand during earthquakes is crucial for improving the flexibility and efficiency of rescue operations. This study proposes anindirect prediction approach involving dimensionality reduction of prediction indicators via kernel principal component analysis, using the resulting principal components as input variables to a long short-term memory (LSTM) neural network. An improved particle swarm optimizationalgorithm was employed to optimize the number of LSTM units and batch size, enabling the development of a casualty prediction model. The predicted casualties were then integrated with safety stock theory to establish an emergency material demand prediction model. Using earthquake data with magnitudes ≥6, the casualty prediction model outperformed others in evaluation metrics such as mean square error, root mean square error, and mean absolute error, reducing errors by 71%—97%, 46%—83%, and 34%—62% compared to the improved GM(1,1), particle swarm optimization-back propagation, and convolutional neural network models, respectively. Case studies of the 2019 M 6.0 earthquake in Changning County (Yibin, Sichuan) and the 2020 M 6.4 earthquake in Payzawat County (Kashgar, Xinjiang) validated the precision of the model in predicting casualties and material demands. The proposed method provides novel technical support for enhancing earthquake disaster management and rescue efficiency, offering remarkable practical value.