AI Foundation Models for Epidemic Intelligence: Integrating Heterogeneous Surveillance Streams through Pre-Training on Historical Outbreak Data

Abstract

Foundation models, pre-trained on massive and diverse datasets and subsequently fine-tuned for specific downstream tasks, have transformed natural language processing, computer vision, and molecular biology. The epidemic intelligence domain, which requires the integration of heterogeneous surveillance streams ; clinical case time series, pathogen genomic sequences, human mobility patterns, environmental surveillance signals, and behavioural and social indicators presents a compelling opportunity for the foundation model paradigm, yet no concrete pre-training methodology for epidemic foundation models has been proposed. This paper introduces the Epidemic Foundation Model, a novel pre-training approach that learns generalised representations of disease dynamics from historical outbreak data spanning more than forty diseases, six decades, and five surveillance modalities, using a federated pre-training architecture that assembles cross-border training data through privacy-preserving gradient aggregation rather than data centralisation. The pre-training methodology employs five complementary objectives: within-stream masked prediction, cross-stream masked prediction, outbreak boundary detection, geographic spread prediction, and cross-disease transfer. The resulting shared model is fine-tunable for outbreak forecasting, variant detection, cross-border alert generation, and intervention decision support. Empirical evidence from deployed multimodal federated surveillance systems demonstrates that integrating all five surveillance streams achieves detection lead times substantially exceeding single-stream baselines, motivating the foundation model approach to learn these cross-stream correlations from historical data rather than re-discovering them from scratch for each new pathogen. The proposed framework is evaluated through leave-one-disease-out cross-validation on forty-two historical outbreaks, demonstrating zero-shot forecasting capability for held-out diseases that substantially outperforms epidemiological model baselines.