Researchers Chenxi Guo and Peter Scott have introduced EpiGIS Pro, an artificial intelligence-powered geospatial intelligence platform designed to unify automated disease event extraction, environmental monitoring, mobility analysis, and predictive epidemiological modeling within a single web-based framework. Global infectious disease surveillance has historically faced operational friction because disparate monitoring streams remain fragmented across distinct operational silos. By consolidating multi-source data ingestion with spatial database infrastructure, EpiGIS Pro establishes a shared data repository encompassing more than 327,000 records across six distinct domains. The system combines structured information extraction driven by large language models, automated scheduled retrieval of environmental and transportation metrics, and machine learning modules capable of forecasting seasonal pathogen surges, detecting statistical anomalies, and evaluating cross-border transmission vulnerabilities. Evaluated across nine countries spanning six years of longitudinal data, the platform provides computational epidemiology with both an operational surveillance instrument and an open, reproducible testbed for evaluating predictive workflows.
The conceptual motivation for the platform stems from persistent limitations in conventional public health surveillance systems, which predominantly rely on laboratory-confirmed diagnostic case notifications that lag behind community transmission trajectories by days or weeks. Event-based surveillance has emerged as a crucial early warning complement by mining unstructured digital sources such as journalistic reporting, official health agency bulletins, and digital communications. Established platforms, including ProMED-mail, HealthMap, and the World Health Organization Epidemic Intelligence from Open Sources initiative, have advanced digital signal aggregation but largely depend on expert manual curation, rule-based extraction, or classical biomedical named entity recognition techniques that struggle with linguistic diversity. Furthermore, existing surveillance systems seldom co-locate environmental determinants of transmission—such as ambient temperature, precipitation, and relative humidity—or human mobility patterns within the same analytical environment. Although spatial analysis techniques such as density clustering, spatial proximity buffering, and disease mapping are central to modern spatial epidemiology, few platforms integrate these capabilities directly with automated predictive analytics.
To overcome these functional boundaries, EpiGIS Pro implements a modular three-tier software architecture comprising a React frontend, a Django REST Framework backend with PostGIS spatial database extensions, and a containerized database layer deployed on Amazon Web Services Elastic Container Service. The platform operates on Python 3.13, Django 6.0, React 19.2, and PostgreSQL 16 equipped with PostGIS 3.4 and the pgvector 0.3 extension for vector similarity operations. Distributed data ingestion and recurring background computations are managed by Celery 5.6 task workers running scheduled Celery Beat routines. Visualization is executed through a dual-stack mapping interface combining the ArcGIS Maps SDK 4.34 with an open-source Leaflet 1.9 and react-leaflet 5.0 fallback layer, ensuring the system can deploy without commercial mapping licenses. The backend is partitioned into twelve dedicated domain applications managing event registries, environmental datasets, transit indices, aviation networks, time-series projections, and healthcare facility geographic directories.
The platform aggregates heterogeneous data streams across distinct geographic scopes, reflecting global coverage for event monitoring and stratified geographic boundaries for localized pilot data. Automated disease event pipelines poll six source categories twice daily or weekly, including NewsAPI, official World Health Organization Disease Outbreak News feeds, Google News RSS streams, GNews, public health subreddits on Reddit, and weekly surveillance bulletins from the Chinese Center for Disease Control and Prevention. Influenza surveillance relies on standardized weekly laboratory reporting from the World Health Organization FluNet platform, compiling 2,868 records across nine countries from December 2019 to February 2026. Auxiliary digital signals include 5,580 weekly Google Trends search interest records for ten health-related queries across six regions. Environmental observations from OpenWeatherMap and Open-Meteo capture 237,277 sub-hourly records spanning eight Australian cities, while the HERE Traffic application programming interface monitors 45 road segments across the same urban centers, contributing 76,334 congestion observations. Aviation mobility data, derived from AviationStack and the OpenSky Network, incorporate 633 active international flight routes connecting 28 monitored airports.
Unstructured event extraction is executed through Anthropic’s Claude 3 Haiku model using structured prompt engineering designed to populate a rigorous epidemiological metadata schema. From ingested media reports, the model extracts pathogen identity, geocoded administrative location, explicit numerical case and death tallies without extrapolation, and a four-tiered severity classification spanning low, medium, high, and critical tiers. The pipeline also captures standardized metadata, including operational priority classifications, structured clinical symptom indicators, zoonotic transmission flags, and clinical syndrome categories such as influenza-like illness, severe acute respiratory infection, gastrointestinal symptoms, and neurological manifestations. Across 353 ingested articles from 56 countries, the language model achieved a 92.4 percent automated processing success rate, averaging 23.5 seconds per document. Core epidemiological fields—including disease classification, severity rating, and priority designation—attained 100 percent field completeness, while quantitative fields exhibited lower completeness rates of 60.1 percent for case counts and 47.5 percent for mortality figures due to incomplete reporting within original journalistic sources.
To facilitate efficient information retrieval and exploratory analysis across large textual databases, EpiGIS Pro incorporates semantic vector search capabilities. Extracted disease event narratives are transformed into 512-dimensional dense vector embeddings using the Voyage AI voyage-3-lite transformer-based embedding architecture. These high-dimensional representations are indexed in PostgreSQL using Hierarchical Navigable Small World graphs via the pgvector extension, facilitating rapid approximate nearest-neighbor spatial queries. This embedding framework empowers public health analysts to execute natural language semantic searches that retrieve epidemiologically relevant events regardless of exact keyword matches. Furthermore, the vector architecture supports automated similar-event retrieval, enabling epidemiologists to identify historical disease incidents that display comparable narrative, geographical, or clinical characteristics to newly emerged outbreaks.
Longitudinal validation of influenza seasonality was conducted using weekly World Health Organization FluNet data spanning seven nations across both geographic hemispheres: Australia, Brazil, South Africa, the United States, Japan, Germany, and France. Seasonal profiles were established by aggregating weekly case observations across 6.2 years to generate mean case counts and standard deviations across the 52 epidemiological weeks. Peak timing was evaluated against established seasonal parameters, defining the austral winter peak between weeks 18 and 39 and the boreal winter peak across weeks 40 through 20. The evaluation demonstrated 100 percent hemisphere-concordant peak timing across all seven evaluated countries. Southern Hemisphere nations peaked between weeks 18 and 27, while Northern Hemisphere nations exhibited peak transmission between weeks 50 and 5. Germany demonstrated the highest peak-to-trough ratio at 457.3, reflecting near-zero inter-seasonal baseline circulation, whereas the United Kingdom and China were excluded from seasonality validation due to data reporting interruptions that precluded consistent multi-year profiling.
Predictive analytics performance was evaluated across four nations possessing continuous reporting volume—Australia, the United States, Japan, and Germany—using a five-fold rolling-origin cross-validation framework with an expanding training window of at least 104 weeks. Three model configurations were benchmarked at forecast horizons of 4, 8, and 12 weeks: an untransformed multiplicative Prophet time-series model with customized ten-order Fourier terms, a conventional log-transformed additive Prophet model, and a seasonal naive persistence baseline that projects future values directly from identical calendar weeks in the preceding year. Multiplicative Prophet achieved positive skill scores exclusively in the United States, where substantial reporting volume and consistent variability enabled the model to achieve a 21 to 29 percent reduction in root mean square error relative to the seasonal naive baseline. Conversely, for Australia, Japan, and Germany, the simple seasonal naive baseline outperformed both complex Prophet architectures. Crucially, applying a logarithmic transformation substantially degraded forecasting accuracy across all nations and horizons, escalating mean absolute percentage error by approximately 110 to 290 percent compared to the multiplicative specification.
The platform pairs forecasting with an ensemble anomaly detection framework and dynamic transmission rate calculations. Anomaly detection integrates three concurrent analytical methods: weekly seasonal Z-scores calibrated against multi-year historical baselines with a default threshold of 2.5 standard deviations, an interquartile range threshold flagging log-transformed values exceeding 1.5 times the interquartile range, and an unsupervised Isolation Forest model parameterized with a 5 percent contamination factor. When tested against documented influenza outbreak seasons in Australia between 2022 and 2024, the conservative consensus requirement of two concurring methods yielded high precision but low individual-week recall, yet achieved a 67 percent event-level recall by correctly detecting two out of three major outbreak seasons. In parallel, the platform estimates the weekly effective reproduction number using a ratio-based formulation incorporating a gamma-distributed serial interval with a three-day mean and a four-week smoothing window. Longitudinal estimation confirmed epidemiologically expected dynamics, with epidemic onset reproduction numbers exceeding trough period figures in Australia (1.27 versus 1.09) and Japan (1.80 versus 1.08), whereas the United States exhibited an anomalous inverted relationship (1.10 versus 1.47) driven by atypical pandemic-era baseline distortions.
Geospatial analysis within EpiGIS Pro executes directly within the spatial database using PostGIS GiST indexing, supporting interactive server-side proximity queries, region containment, and cluster detection. Distance-based proximity searches compute geodesic buffers using spatial predicates to rank surrounding disease events within user-specified radii. Spatial clustering relies on density-based spatial clustering of applications with noise via the ST_ClusterDBSCAN function, grouping disease points within a default neighborhood distance of 300 kilometers and a three-point minimum threshold, while computing convex hulls and centroids for detected spatial clusters. Per-country spatial summaries generate bounding hulls and spread radii color-coded by event frequency. To quantify cross-border vulnerability, the system links active departure-country disease occurrences with 633 international aviation corridors, computing automated daily risk scores scaled from zero to 100 based on active case counts, disease severity, and estimated daily passenger flows across 4,828 corridor assessments.
Notwithstanding its technical breadth, the authors delineate several critical methodological boundaries and design limitations. The disease event extraction corpus of 353 articles remains modest in scale, reflecting early operational deployment rather than an exhaustive historical repository. Forecasting accuracy demonstrates pronounced sensitivity to regional data volume, reaffirming that statistical baselines frequently surpass parameterized time-series models in sparse reporting environments. The anomaly detection module’s low individual-week recall underscores the persistent difficulty of distinguishing true epidemiological surges from normal seasonal variance when annual amplitudes fluctuate widely. Additionally, cross-border aviation risk scoring currently utilizes empirical multiplicative equations rather than mechanistic susceptible-exposed-infectious-recovered compartmental transmission modeling. Crucially, the environmental weather observations and urban traffic indicators are strictly confined to an Australian pilot and function solely as contextual visualization layers; they are not incorporated as predictive covariates in the reported time-series models. Planned developments include expanding environmental data collection internationally, integrating weather and transit covariates as exogenous regressors, implementing mechanistic compartmental transmission models, and testing the system within public health agency workflows.
Subject of Research: Integrated geospatial disease surveillance and predictive analytics platform
Article Title: EpiGIS pro: an AI-powered geospatial intelligence platform for integrated disease surveillance and predictive analytics
Article References: Guo, C., & Scott, P. (2026). EpiGIS pro: an AI-powered geospatial intelligence platform for integrated disease surveillance and predictive analytics. International Journal of Health Geographics, 25(1), Article 45. https://doi.org/10.1186/s12942-026-00479-1
Image Credits: AI Generated
DOI: 10.1186/s12942-026-00479-1
Keywords: infectious disease surveillance, geospatial intelligence, EpiGIS Pro, large language models, Claude AI, PostGIS, Prophet forecasting, WHO FluNet, effective reproduction number, aviation mobility, anomaly detection, spatial epidemiology, vector embeddings, public health informatics, time-series modeling, event-based surveillance
News Source: Blake Davidson. (November 19, 2026). EpiGIS Pro Integrates AI Disease Tracking With Mobility and Predictive Analytics. Scienmag.



