LOVS: turning public disease outbreak data into spatial maps, forecasts, and searchable intelligence

The problem

When an infectious disease outbreak occurs in a complex humanitarian setting, the response typically involves many different international and local actors, including government health authorities, WHO, UNICEF, international NGOs, and local organisations. Each of these actors collects and publishes data, evidence, and operational information, but this material is scattered across different channels and organisations. It is often shared publicly but reaches other responders through informal routes.

The challenge is that no single organisation has a consolidated view of the entire body of evidence. A responder working in one health zone may not have visibility into what is happening in neighbouring zones, what other organisations are reporting, or how different factors are interacting to shape the trajectory of the outbreak.

In eastern Democratic Republic of the Congo (DRC), where the 2026 outbreak of Ebola disease caused by Bundibugyo virus (BDBV) emerged, the response involves a large number of operators working across a wide and difficult-to-access area. The region faces long-standing humanitarian challenges, and the response operates under significant constraints including restricted access to some health zones, high demand on healthcare facilities, and complex security conditions.

The solution

LOVS, the Latent Outbreak Visibility System, is an intelligence and epidemiological tool that brings together publicly available data and evidence from across the many different actors responding to the Ebola outbreak, synthesises it into a single structured format, and makes it available for analyses by both humans and AI.

The tool draws on data and evidence published by the DRC Ministry of Health, provincial and health zone public health authorities, WHO, UNICEF, international NGOs, and other actors. It collects operational evidence, statistical evidence, and situational reporting that these organisations share publicly, and brings it into one corpus. The evidence corpus, updated nearly daily, contains over 7,000 different cited evidence claims and data points across 305 evidence assets, from confirmed case counts to calculations such as child case fatality ratios.

LOVS has three main features.

• The first is a spatial evidence map, which provides a visual overview of where the outbreak currently stands. It models the percentage probability of the outbreak spreading to neighbouring health zones and identifies corridors where it might reach a major city or cross into another country. These corridors are pre-committed watch points for attention, not necessarily predictions of spread.

• The second is a natural-language interface that allows users to ask questions against the entire evidence corpus. A journalist, responder, or decision-maker can ask about a specific health zone and receive cited insights drawn from the underlying evidence. For example, the tool can surface that a particular health zone has had a stagnant confirmed caseload, which could possibly be attributed to another evidence point, such as access to the area has been restricted. This helps with modelling and inference, pointing toward a likely higher real caseload in that zone that is currently less visible to the response. Every insight is cited back to a primary source.

• The third is a methodological brief containing epidemiological calculations including burden nowcasting, ascertainment and visibility estimates of the real caseload and death toll, and charts tracking daily changes. It explains the statistical methods used, and it carries a calibration thread that records forecasts and scores them publicly in the future. Modelled estimates are transparent and reproducible, and the confirmed counts are treated as a floor.

The tool was originally prototyped as a system for predicting when infectious disease outbreaks might emerge, based on publicly available data points including academic articles, genomic datasets, weather, deforestation, climate data, displacement events, and population movements. On 20 April 2026, the prototype predicted that there would be, within the next six months at least one new international laboratory-confirmed filovirus spillover event in sub-Saharan Africa. The Bundibugyo Ebola outbreak was confirmed in the DRC on 14 May 2026, twenty-four days after that forecast.

When the outbreak was confirmed, the prototype was operationalised in approximately one week. AI coding tools including Anthropic's Claude and OpenAI's Codex were used alongside manual coding, and all statistical models and code were reviewed and validated given the sensitivity of the subject matter.

The tool is maintained on a daily cadence. This is made possible by an internal technology stack previously built to coordinate and orchestrate AI agents. This same infrastructure enables the ongoing expansion of LOVS as new analytical needs emerge. For example, analyses of the data led to the addition of vulnerability modelling for specific communities, using multisectoral data on water and sanitation, food security, malaria burden, and healthcare access, developed with the Lake Tanganyika Floating Health Clinic. Work is underway to model broader health service impacts in areas where outbreak response is absorbing and exceeding significant healthcare capacity.

LOVS uses only publicly available data that does not contain personal information. This was a deliberate design choice to ensure the tool could be shared openly and to facilitate collaboration. It was created to support the responding authorities, and does not speak for any of the institutions cited.

The impact

1. Dated forecasting resolved against official records

At set intervals when the operational reality shifts, forecasts with fixed resolution dates are published to publicly showcase a potential forecast and determine whether it is correct. For example, on 22 June 2026, multiple forecasts were published that were resolved on 19 July against official data, which included: (a) a prediction that national confirmed cases would be >1,965, which resolved at 2,423 from official sources, and (b) a prediction that national confirmed deaths would be upwards of 988, which resolved at 967 from official sources.

2. Insights shared with UNICEF informed their updated response plan

On 12 June 2026, a deep analysis was run against the LOVS evidence corpus using large language models, producing a series of insights, analyses, and recommendations. These were shared with UNICEF. A month later, at the end of July 2026, UNICEF updated its outbreak response plan in alignment with those insights and analyses after careful consideration and human review. LOVS enabled insights to be drawn out from scattered operational data to identify blind spots and ways in which the response could be strengthened, in a time- and cost-efficient way.

3. Analysis surfaced coordination gaps that proved significant

The tool identified gaps in coordination between responding organisations in specific areas. In operational emergencies with multiple actors, these are known as "seams," points where two responders have overlapping responsibilities, but neither is fully accountable for the handover between them. LOVS surfaced breakdowns in these seams, particularly highlighting the transition between contact tracing and patient care as well as areas where conflict and access restrictions were limiting the response. These gaps were flagged when the affected area had a relatively small caseload. Given that this is a live outbreak where LOVS is being applied, conditions in any area can change significantly in seemingly unpredictable ways. For example, from mid-June to late July 2026, there was a notable increase in cases and pressure on healthcare facilities in one province, suggesting that the on-demand AI analysis of the live outbreak had identified a meaningful signal that took a month to become clearly visible in the official confirmed numbers.

4. Cross-operator analysis at a scale not previously available

At the time of writing, LOVS enabled analysis across approximately 7,227 cited evidence claims from official sources and 137 different operator activities in response to the outbreak, identifying where coordination between organisations could be strengthened based on their public reporting. This type of analysis across the full evidence corpus, spanning multiple actors and data sources, would not have been practical to conduct manually.

Lessons learned

• AI gives a single individual with deep expertise extraordinary leverage. The tool was built and is maintained by one person, using AI coding agents working in parallel. During the initial sprint to operationalise the tool, up to 10 coding agents were used around the clock. This kind of leverage, the ability to bring intelligence on demand without being limited by team size, is fundamentally new and relevant for any organisation working in time-sensitive contexts.

• Internal tooling and systems matter as much as the end product. Before building a system like LOVS, organisations should think about what internal tooling and technology they need to amplify their ability to leverage AI effectively. Building an internal technology stack enabled coordination and orchestration of AI agents, which is also how LOVS is maintained daily.

• Using only publicly available data enables openness and collaboration. The decision to use only publicly available, non-personal data meant the tool could be shared openly, discussed with multiple organisations, and developed without privacy constraints. However, more granular data would enable finer-scale and more contextual analysis.

• Review remains essential for sensitive applications. Despite the speed and scale that AI coding tools provide, all code and statistical models are reviewed because of the sensitivity of the subject matter. For applications where the outputs inform decisions in a crisis context, human review of both the code and the analytical outputs is essential. This is commonly referred to as keeping a ‘human in the loop.

• The long-term vision is prediction, not just response. LOVS in its current form supports response to an active outbreak. The longer-term goal is a system that monitors health zones continuously, using primary care data alongside publicly available data on weather, deforestation, climate, displacement, and population movements, to predict when an outbreak is beginning to emerge and signal it in near-real time.