关键词: 突发公共卫生事件, 时空注意力机制, 时空序列依赖性, 应急医疗物资需求预测

Abstract:

The outbreak of public health emergencies， such as COVID-19， necessitates accurate predictions of epidemic trends and medical resource demands to enable timely interventions. Existing methods for epidemic forecasting often fail to simultaneously capture spatial dependencies and temporal dependencies. Additionally， dynamic demand prediction for emergency medical supplies， classified as consumable supplies and non-consumable supplies， remains to be unexplored， particularly under rapidly evolving scenarios. Therefore， it tries to address these gaps for this study by proposing a novel framework for spatiotemporal epidemic prediction and resource demand modeling. A SpacetimeNet is introduced， which is an autoregressive model integrating spatiotemporal attention mechanisms to jointly learn spatial correlations and temporal dependencies. The framework comprises： （1） Spatial Encoder： It employs multi-head self-attention to model dynamic inter-regional influences. For target region x_t^i(i\in N,t\in N)， the spatial attention vector Attentio{n}_i(Q_i,K,V) is computed by Attentio{n}_i(Q_i,K,V)=softmax(Q_{\mathrm{i}}{K}^T/\sqrt[]{d_k})V. （2）Temporal Encoder： It combines positional encoding with temporal attention to capture sequential patterns. （3）Autoregressive Decoder： It uses masked self-attention to iteratively predict future values while preventing information leakage. The final prediction {\widehat{y}}_t^i is derived via {\widehat{y}}_t^i=ReLu(W^p{O}_i^{\text{'}}+b^p).The used dataset includes 13 features （e.g.， confirmed cases， deaths， vaccination rates） from 5 representative countries （France， Germany， India， USA and UK）. Data preprocessing involves log-transformation and normalization to enhance stationarity. An 80：20 train-test split ensures robust evaluation. Daily new cases are forecasted using the SpacetimeNet， and then actual active cases are computed and medical resource demands are derived based on mortality/recovery rates and staffing ratios.Evaluated on the dataset described， the SpacetimeNet outperformed benchmarks （ARIMA， LSTM， GRU， etc.） with lower MAE， lower RMSE and lower SMAPE. The final results show that in the early stage of the rapid development of the epidemic， the demand for consumable emergency medical supplies rises rapidly， while the demand for non-consumable emergency medical supplies fluctuates considerably， after which the rise in the demand for consumable emergency medical supplies gradually levels off， and the demand for non-consumable emergency medical supplies gradually decreases due to the accumulation of resources to 0. The results of the present study will provide a precise prevention and control of epidemics and the allocation of resources in the context of a public health emergency.

Key words: public health emergencies, spatiotemporal attention mechanism, spatiotemporal sequence dependence, emergency medical supplies demand forecasting