💻 computer science

Predicting Pathogen Emergence Using a Spatiotemporal Graph Neural Network and Explainable AI Approach

This study proposes a framework integrating Spatiotemporal Graph Neural Networks with Explainable AI to accurately forecast pathogen emergence risks using a human epidemic database, achieving superior performance over baseline models while quantifying uncertainty and identifying key drivers to enable proactive public health interventions.

Original authors: Oreofe Jolaolu, Okoh J.C

Published 2026-08-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Oreofe Jolaolu, Okoh J.C

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

The world of infectious disease forecasting has long been a battleground between two competing philosophies. On one side stand the traditional models, which rely on clear, step-by-step rules about how diseases spread, much like a map with fixed roads. These are easy to understand but often fail when the terrain changes unexpectedly. On the other side are the powerful, modern machine learning tools that can spot hidden patterns in massive amounts of data, but they often operate like a black box, offering predictions without explaining how they reached them. For public health officials, this lack of transparency is a major problem; they need to know not just that an outbreak is coming, but why, so they can act with confidence. The challenge has been to build a system that combines the predictive power of advanced computers with the clarity needed for human decision-making, all while accounting for the fact that diseases do not respect borders and move through space and time in complex, shifting ways.

A recent study by researchers Oreofe Jolaolu and Dr. Okoh J.C. tackles this exact problem by creating a new framework that merges these two worlds. They developed a system that treats the globe as a connected network, where every country or region is a point on a map, and the lines connecting them represent the ways people and pathogens travel between them. This structure, known as a spatiotemporal graph, allows the computer to see how an outbreak in one place might ripple out to its neighbors over time. By feeding this system a massive collection of historical data—specifically, 1,044 recorded epidemic events from 2015 to 2020 involving over 120 different pathogens—the researchers trained the model to learn the intricate dance of disease emergence. Unlike older methods that looked at time or space separately, this approach watches both simultaneously, capturing how a virus moves across a border while also evolving day by day.

The results of this experiment were striking. When the researchers tested their new system against standard forecasting tools, their model proved significantly more accurate. It achieved a score of 0.6959 in predicting future outbreak trends, a substantial improvement over the next best method, which scored only 0.5685. The older, simpler models that looked only at time or only at space performed poorly, with one even failing to predict anything better than random chance. This confirmed that to understand disease spread, one cannot simply look at a timeline or a map in isolation; the two must be woven together. The model successfully identified that the most important factors driving these predictions were the number of past cases, the total number of deaths, and the specific type of pathogen involved. These findings suggest that the system is not just guessing but is learning the genuine, underlying mechanics of how outbreaks grow and move.

Perhaps the most significant contribution of this work is that it does not hide its reasoning. In a field where trust is paramount, the researchers equipped their model with a set of tools designed to explain its own thinking. By using a method that breaks down every prediction into its contributing parts, they showed exactly which data points were pushing the model toward a specific warning. This transparency means that health officials can see the logic behind a forecast, such as a spike in risk driven by a specific pathogen's history in a region, rather than receiving a mysterious alert. Furthermore, the study went a step further to measure how much of the uncertainty in the predictions came from the data itself versus the model's own confusion. They found that the vast majority of the uncertainty, nearly 99 percent, came from the natural randomness of the data—the inherent noise of real-world events—while the model itself was remarkably stable and confident in its calculations.

This research offers a promising path forward for global health security. By combining a network that understands how places connect with a system that explains its own logic, the framework provides a tool that is both powerful and trustworthy. It moves the field away from passive observation, where officials wait for an outbreak to happen, toward proactive management, where resources can be directed to high-risk areas before a crisis escalates. While the current model relies on static connections between regions, the researchers note that future versions could adapt to changing human movements, such as travel restrictions or seasonal shifts. For now, this work stands as a clear demonstration that it is possible to build artificial intelligence that not only predicts the future of infectious diseases but also tells us exactly why it thinks so, laying a vital foundation for a safer, more prepared world.

Technical Summary: Predicting Pathogen Emergence Using a Spatiotemporal Graph Neural Network and Explainable AI Approach

Problem Statement

Traditional epidemiological forecasting faces significant challenges due to the complex spatiotemporal dynamics of emerging infectious diseases, exacerbated by climate change and global mobility. Existing approaches fall into three categories with distinct limitations:

Compartmental Models: Offer mechanistic clarity but struggle with high-dimensional data (e.g., climate, mobility) and assume homogeneous mixing.
Statistical Models: Effective for baseline time-series analysis but lack generalizability for complex, multimodal digital surveillance data.
Machine Learning (ML): Capable of discovering nonlinear patterns but often operate as "black boxes," lacking the transparency required for One Health decision-making where diverse stakeholders need actionable rationales.

Furthermore, current Spatiotemporal Graph Neural Network (STGNN) applications often rely on oversimplified spatial assumptions (treating locations with similar features as having identical future trajectories) and neglect underlying mechanistic heterogeneity. There is a critical gap in frameworks that integrate epidemiology-informed modeling, spatiotemporal graphs, and Explainable AI (xAI) to provide both accurate global forecasting and transparent, actionable insights.

Methodology

Data Source and Preprocessing

The study utilizes the Human Epidemic Database (HED), a structured dataset derived from official, open-source surveillance reports.

Scope: 1,044 epidemic events across various countries and territories.
Timeframe: Disease onset between 2015 and 2020.
Coverage: Over 120 distinct pathogens.
Feature Engineering: The raw data was transformed into a multi-dimensional tensor with dimensions [Batch, Nodes, Timesteps, Features].
- Nodes (N): Discrete spatial regions/countries (ISO identifiers).
- Temporal Horizon (T): 225 timesteps of historical lookback.
- Features (F): Cumulative cases ( $F_0$ ), cumulative deaths ( $F_1$ ), and pathogen taxonomy groupings ( $F_2$ ).
- Graph Structure: A symmetric spatial adjacency matrix ( $A$ ) models connectivity between nodes.

Model Architecture: Spatiotemporal Graph Neural Network (STGNN)

The proposed STGNN jointly processes spatial graph topologies and temporal dynamics:

Spatial Graph Convolution: Aggregates localized structural information, allowing neighboring nodes to share contextual signals via the adjacency matrix.
Temporal Sequence Layers: Processes longitudinal trends using hidden layers with 128 units.
Poisson Rate Head: The network terminates in a Poisson Rate Head ( $\lambda$ ) to model discrete epidemiological event counts.
Output Activation: Utilizes a Softplus activation function to output the Poisson rate parameter ( $\lambda$ ), representing the expected mean count per node-snapshot.
Training: Optimized using the Adam optimizer (learning rate 0.001, L2 decay 0.0001) with a batch size of 32. The loss function is Poisson Negative Log-Likelihood (NLL). Training runs for up to 100 epochs with early stopping (patience = 10).

Explainable AI (xAI) and Uncertainty Quantification

To address the "black box" limitation, the framework integrates:

Feature Attribution: Uses Kernel SHAP and LIME to extract global feature importance and local, prediction-specific explanations. A dimensional unflattening pipeline maps flat explanation vectors back to structural coordinates (Node Index, Time Lag, Feature Name).
Uncertainty Quantification (UQ): Employs Monte Carlo Dropout (50 stochastic passes, 20% dropout rate) to decompose predictive variance into:
- Aleatoric Uncertainty: Inherent data noise.
- Epistemic Uncertainty: Model parameter ambiguity.

Baseline Comparisons

The STGNN was benchmarked against:

Standard GCN: Spatial propagation without temporal sequence lookbacks.
Standard LSTM: Longitudinal patterns without spatial graph structures.
XGBoost: Non-deep learning baseline on flat feature representations.

Key Results

Predictive Performance

The proposed STGNN significantly outperformed all baseline architectures across rolling temporal test sets:

Proposed STGNN: $R^2 = 0.6959$ , MAE = 0.4694, RMSE = 1.2385.
XGBoost: $R^2 = 0.5685$ , MAE = 0.5583.
Standard LSTM: $R^2 = 0.1587$ , MAE = 0.8171.
Standard GCN: $R^2 = -0.0423$ , MAE = 0.9495.

The results confirm that isolated spatial (GCN) or temporal (LSTM) models fail to capture the multi-dimensional nature of epidemiological data, while the STGNN's joint modeling is essential for capturing cross-regional transmission pathways.

Explainability Findings

SHAP analysis identified the primary drivers of the model's output:

Cumulative Cases ( $F_0$ ): The primary global driver, reflecting standard autoregressive behavior where past realizations condition future trajectories.
Cumulative Deaths ( $F_1$ ): A secondary severity weighting factor anchoring outbreak intensity.
Pathogen Group Taxonomy ( $F_2$ ): A structural control allowing the network to condition propagation rates based on pathogen class.

Uncertainty Quantification

Analysis of 944 validation instances revealed:

Aleatoric Dominance: 98.85% of total predictive variance is attributed to inherent data noise (consistent with Poisson distributions where mean equals variance).
Epistemic Convergence: Only 1.15% of variance is due to model parameter ignorance, indicating high network convergence and weight stability.
Calibration: A strong positive correlation ( $r = 0.6101$ ) exists between predicted standard deviation and absolute prediction error, confirming the model reliably signals high-risk anomalies.

Significance and Contributions

The paper claims the following contributions to the field of One Health and disease forecasting:

Superior Forecasting Accuracy: Demonstrates that integrating spatiotemporal graph messaging is necessary to outperform traditional ML and isolated deep learning models.
Transparency in AI: Validates that xAI techniques (SHAP/LIME) can link model predictions to logical clinical drivers, fostering trust among public health professionals.
Operational Risk Mitigation: The framework provides a method to quantify uncertainty, showing that predicted standard deviation can serve as a dependable early-warning indicator for high-risk spatial-temporal anomalies.
Foundation for Future Integration: The study lays a foundation for integrating broader climate data variables and dynamic adjacency matrices (to account for changing human mobility) in future iterations.

The authors conclude that this framework empowers global health authorities to transition from passive surveillance to proactive, data-driven resource allocation, strengthening health security against future pathogenic threats.