Optimizing spatiotemporal nonpharmaceutical interventions for influenza: An adaptive reinforcement learning approach for regional heterogeneity

Background

Influenza remains a significant global public health challenge because of its high transmissibility, widespread circulation, and considerable societal impact. Conventional threshold-based nonpharmaceutical interventions (NPIs) provide valuable frameworks for outbreak control; however, these standardized approaches may not fully account for important regional heterogeneity. It remains difficult to weigh regional characteristics and accurately balance infection control and socioeconomic costs.

Methods

We propose a susceptible-exposed-infectious-quarantined-removed compartmental model-dueling deep Q-network (SEIQR-Dueling DQN) framework tailored to plains, hilly, and plateau cities. By integrating climatic, demographic, and health care resource data, the model captures regional differences in transmission and recovery dynamics. A multidimensional state space and discrete intervention set allow for the adaptive optimization of NPI strategies across varying epidemic and resource conditions. Model parameters were estimated by sequential Bayesian optimization, and bootstrap resampling was used to quantify uncertainty. In addition, the performance of the SEIQR-Dueling DQN strategy was compared with that of the threshold-based strategy in terms of the reduction in cumulative infections, peak prevalence and length of intervention periods.

Results

The threshold-based intervention policy reduced cumulative infections by 3.05 %–3.67 % and peak incidence by 8.26 %–12.58 % but showed limited responsiveness to regional variation, often resulting in either under- or over-control. The SEIQR-Dueling DQN framework dynamically adjusted intervention timing and combinations on the basis of local demographic structures and epidemic trends and reduced cumulative infections by 5.87 %, 5.99 %, and 5.21 % in plains, hilly, and plateau cities, respectively, while achieving peak reductions of 34.92 %, 22.23 %, and 8.12 %, respectively, with a balanced consideration of socioeconomic impact. To assess generalizability, the trained model was applied to cities with differing transmission dynamics and demonstrated consistent performance across settings.

Conclusion

The SEIQR-Dueling DQN framework supports tailored interventions across regions and shows promise for broader application in the management of regional heterogeneity and future emerging infectious diseases.