When a novel pathogen emerges, public health officials, clinicians, and machine learning researchers face a critical "cold start" problem. Designing predictive models for patient mortality, regional transmission, or therapeutic discovery typically requires massive amounts of historical, disease-specific data. In the opening weeks of an outbreak, that data simply does not exist. Waiting for clinical registries to mature or case counts to accumulate costs lives.
To bypass this data lag, a powerful paradigm has emerged: epidemiological transfer learning. Instead of training models from scratch on a blank slate, researchers leverage decades of deep, historical data from existing diseases to predict the behavior of a novel threat.
Why Knowledge Transfers: The Invariance of Host and Environment
At first glance, transferring knowledge across distinct pathogens seems risky. A new virus might have a different incubation period, a higher fatality rate, or unique molecular targets. Why should a model trained on seasonal influenza or general respiratory infections work for a novel coronavirus?
The core intuition relies on a profound realization: while the pathogen is new, the host and the environment are not.
- Conserved Physiological Pathways: The human body has a finite number of ways to respond to severe infection. Complications like acute respiratory distress syndrome (ARDS), sepsis, or hyperinflammatory cytokine storms share common biological underpinnings regardless of the triggering agent.
- Stable Geospatial Plumbing: A respiratory virus spreads through human contact. The underlying social networks, commuting patterns, school calendars, and neighborhood densities of a region—what we can think of as the "geospatial plumbing"—remain completely stable over time.
- Universal Grammar of Contagion: Epidemic time-series are governed by fundamental physical and mathematical constraints. Concepts like exponential growth, seasonal periodicity, and the dampening effects of public health interventions follow universal mathematical laws.
Clinical Risk and Outcome Prediction
How do we build a personalized risk calculator for a disease we have never seen before? The solution lies in training models on proxy outcomes from well-documented conditions.
Williams et al. demonstrated that models trained entirely on historical influenza data can serve as highly accurate proxies for predicting COVID-19 hospitalizations and deaths. Because demographic vulnerabilities (like age and baseline comorbidities) are highly transportable between symptomatically similar respiratory conditions, the model's patient risk rankings remain robust.
Similarly, Barda et al. addressed the complete absence of individual-level data by training a baseline model on general respiratory infections, then using aggregate, macro-level death rates from early-hit regions to recalibrate the model. This allowed them to deploy a high-performing mortality risk model at the very start of a pandemic.
This transferability extends deep into intensive care dynamics. Lichtner et al. showed that a machine learning model trained on non-COVID-19 viral pneumonia patients could effectively predict lethal outcomes in critically ill COVID-19 patients, outperforming standard clinical scores by capturing the shared physiological mechanisms of organ failure.
To handle complex, multi-modal clinical data, Agarwal et al. introduced TransMED, a hierarchical framework that pre-trains on large cohorts of Severe Respiratory Disease (SRD) patients. By exploiting the high overlap in medical vocabulary and clinical patterns, the model successfully predicts mechanical ventilation needs for emerging pathogens.
Remarkably, this biological commonality allows knowledge to flow in reverse as well. Zhao et al. utilized data from COVID-19 patients to pre-train a model that predicts severe cytokine release syndrome (CRS) in cancer patients undergoing CAR-T cell therapy, proving that systemic inflammatory insights are highly fungible across entirely different medical domains.
Geospatial Transmission and Regional Forecasting
Predicting where a virus will strike next requires mapping how populations interact with their local environment. Instead of waiting for local outbreaks to flare up, researchers can read a region's "risk phenotype" from its historical data.
Chattopadhyay formalized this by analyzing a decade of historical seasonal flu data through information theory to compute a "Universal Influenza-like Transmission" (UnIT) score for US counties. Because the structural factors facilitating spread are invariant, this flu-derived score significantly improved localized COVID-19 incidence forecasts.
When an outbreak begins in a specific epicenter, cross-city transfer learning can protect uninfected regions. Xiao et al. developed C-Watcher, which uses an adversarial encoder-decoder to extract "city-invariant" mobility patterns from early epicenters. By transferring these behavioral signatures to target cities, the framework identifies high-risk neighborhoods before the first local cases are even clinically detected.
Furthermore, the structural dynamics of public health interventions can be transferred across borders. Hu et al. applied a modified auto-encoder to transfer learned containment dynamics from mature outbreak regions to areas in the earliest stages of a pandemic, allowing local authorities to simulate the impact of rigorous interventions before implementing them.
Epidemic Foundation Models and Zero-Shot Adaptability
The ultimate realization of this paradigm is the shift away from single-pathogen models toward epidemic foundation models. Rather than treating every outbreak as an isolated event, these architectures treat all epidemics as variations of a shared underlying grammar.
Kamarthi et al. introduced PEMS (Pre-trained Epidemic Time-series Models), a framework that uses self-supervised learning on over a century of historical data spanning diverse diseases. By learning universal contagion patterns and seasonal periodicities, PEMS achieves zero-shot adaptability, outperforming specialized baselines on entirely new pathogens without requiring disease-specific training.
Looking forward, Metcalf et al. outline the broader path toward AI foundation models for public health. They highlight how large models pre-trained on diverse historical outbreak data can learn generalizable principles of disease dynamics across different pathogens and populations, providing rapid, adaptable insights that are especially crucial for resource-constrained environments facing novel biological threats.
Even when tracking existing diseases through the noise of a new crisis, historical models remain vital. Rodríguez et al. developed CALI-NET to "steer" historical flu forecasting models through the data anomalies caused by COVID-19, demonstrating that existing epidemiological expertise can be dynamically adapted rather than discarded. This universal perspective is mirrored in non-parametric approaches like EpiLearn, developed by Morel et al., which forecasts local incidence curves by searching a global database for "historical twins"—past trajectories from entirely different regions and times that match the current mathematical progression.
Accelerating Therapeutics and Information Retrieval
The utility of cross-disease transfer extends beyond forecasting into molecular biology and information management. In drug discovery, the challenge is identifying effective treatments when specific screening data is non-existent.
Stokes et al. developed ComboNet, a deep learning architecture that predicts synergistic drug combinations for novel pathogens by leveraging drug combination data from taxonomically or biologically related diseases, such as HIV. By exploiting shared biological interactome similarities, the model learns robust representations that generalize to new viral threats. To support these efforts, Arshadi et al. introduced CoronaDB-AI, a curated dataset designed to train deep learning models by leveraging data from taxonomically related viruses, mitigating initial data scarcity for in silico screening and vaccine design.
Finally, even the task of navigating the sudden explosion of scientific literature during a crisis benefits from cross-domain transfer. Cohan et al. demonstrated with SLEDGE-Z that relevance signals learned from general medical information retrieval tasks transfer seamlessly to novel disease literature, allowing search engines to surface critical clinical insights without requiring specific training on the emerging pathogen.
Key Foundational Work
Predicting lethal courses in critically ill COVID-19 patients using a machine learning model trained on patients with non-COVID-19 viral pneumonia (2021) — contributed 3 concept elements to this direction.
Developing a COVID-19 mortality risk prediction model when individual-level data are not available (2020) — contributed 3 concept elements to this direction.
Forecasting and Evaluating Multiple Interventions for COVID-19 Worldwide (2020) — contributed 2 concept elements to this direction.