BioSurveillance-Agent: A Multi-Method Explainable Agentic Platform for Epidemic Surveillance on One-Health Networks
BioSurveillance-Agent is a multi-method explainable agentic platform that integrates One Health networks, wearable sensors, and a hybrid AI pipeline to enable context-adaptive, hierarchical epidemic surveillance and intervention, demonstrating significant reductions in disease spread and mortality through robust, human-auditable decision-making.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
When a new disease begins to spread, the race is not just to find it, but to understand how it moves through the complex web of human, animal, and environmental connections that make up our world. This is the domain of "One Health," a perspective that recognizes that the health of people is deeply tied to the health of livestock, wildlife, and the environment around them. In this field, scientists build digital models to simulate how infections travel, much like a weather forecast predicts a storm. However, a critical challenge remains: if a computer program decides to lock down a city or distribute vaccines, public health officials need to know exactly why the computer made that choice. They cannot act on a "black box" decision that offers no explanation. The goal is to create a system that is not only fast and accurate but also transparent, allowing humans to audit and trust the machine's reasoning before they act.
Researchers at the University of Catanzaro and the University of Calabria have built a new platform called BioSurveillance-Agent to solve this problem. It is a unified system designed to watch for epidemics, reason about how to stop them, and deploy countermeasures, all while keeping a human in the loop to verify the logic. The platform simulates a world where people, animals, and the environment are connected in a vast network. It uses data from simulated wearable sensors—devices that track heart rate, body temperature, and activity—to detect early signs of illness in individuals before they even show obvious symptoms. These signals are fed into a sophisticated computer model that predicts who is likely to get sick next. But the true innovation lies in how the system explains its predictions. Instead of relying on a single method to justify its decisions, the agent asks five different types of interpretable models and four advanced mathematical tools to analyze the same data. It then combines their answers to create a clear, ranked list of who is most at risk and why.
The system operates like a team of digital agents working together. In a more advanced setup, a central coordinator manages several regional agents, each responsible for a specific area. These regional agents are not identical; they are given different "capability profiles" that reflect the reality of their local environment. A rural region with limited vaccine stock and no mandate for quarantines will act differently than an urban center with full hospital capacity and strict isolation rules. The central agent watches the big picture, stepping in only when a region runs out of resources or when the disease spreads across borders. This design allows the system to adapt its strategy to the specific constraints of each location, rather than applying a one-size-fits-all rule.
To test how well this approach works, the researchers ran a series of rigorous simulations. They compared the new multi-method system against older, simpler approaches and found that the hybrid design was highly effective. The system successfully reduced the overall spread of the epidemic and lowered the number of deaths compared to scenarios where no intervention occurred. Specifically, the machine-guided interventions reduced the total burden of the epidemic by nearly a quarter and cut the death toll by more than 13 percent. When the researchers compared a single, flat agent managing the whole network against the hierarchical team of regional agents, the team approach proved superior. The regional agents deployed nearly three times as many targeted interventions because they could act simultaneously in their own areas, rather than waiting for a single central brain to process every alert.
A key part of the research was determining which method of explanation was best for human understanding. The team tested nine different ways to explain the computer's decisions. They found a clear divide between two types of methods. The first type, known as classical interpretable models, produced explanations that looked like simple "if-then" rules, such as "if the heart rate is high and activity is low, then the person is at risk." These were easy for humans to read and understand, scoring very high on clarity. The second type, known as post-hoc methods, used complex mathematical gradients to explain the decision. While these were mathematically sophisticated, they produced explanations that looked like lists of numbers with no clear meaning to a human reader, such as "embedding dimension 13 is the cause." The study showed that the system could achieve high accuracy using the clear, rule-based explanations without sacrificing performance, challenging the old belief that you must choose between a smart computer and a clear explanation.
The researchers also tested the system's resilience by deliberately breaking parts of it. They added random noise to the data, hid key information, and even removed entire explanation methods to see if the system would fail. Surprisingly, the system's ability to reduce deaths and control the epidemic remained steady even when the explanations were scrambled. This happened because the system's decision-making logic was built on simple thresholds—like "if the infection rate goes above a certain point, start quarantining"—rather than relying on the precise details of any single explanation. This means the system is robust; it can handle errors in its reasoning tools without collapsing, because the overall strategy is grounded in solid, pre-defined safety limits.
Finally, the team simulated a world where resources changed in the middle of an outbreak. They tested scenarios where a region suddenly ran out of vaccines or received a fresh shipment. The agents adapted immediately. When a region's supplies were cut, the local agent stopped acting and asked the central coordinator for help. When supplies were replenished, the agent resumed its work. This dynamic adaptation ensured that resources were used where they were needed most, and it reduced the number of deaths further than a static system could. The study concludes that by combining a network of smart, adaptable agents with a multi-layered approach to explaining their decisions, it is possible to build an epidemic surveillance system that is both powerful enough to save lives and transparent enough for humans to trust.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.