Epidemiology-Informed Graph Neural Networks for Predicting and Interpreting Transmissible Hospital-Acquired Infections: A Retrospective Cohort and Simulation Study

This paper proposes an epidemiology-informed graph neural network (EIGNN) framework that integrates mechanistic epidemiological models with data-driven contact networks to accurately predict and interpret hospital-acquired infection dynamics while ensuring clinical trust through transparency.

The Gist

Imagine a hospital as a bustling, ever-changing city. In this city, people (patients) and workers (doctors, nurses) move between different neighborhoods (wards and rooms). Sometimes, a "virus" acts like a mischievous ghost that jumps from person to person when they get too close.

Predicting where this ghost will go next is incredibly hard. It's not just about who is standing next to whom right now; it's about who was there yesterday, who is likely to get sick tomorrow, and how the "ghost" behaves based on biological rules (like how long it takes to recover).

For a long time, computer scientists have tried to solve this using Graph Neural Networks (GNNs). Think of these as super-smart detectives who look at the map of the city and the movement of people to guess who gets sick next. They are great at spotting patterns, but they are like "black boxes." You ask them, "Who will get sick?" and they say, "This person," but they can't explain why based on the rules of biology. They just guess based on data, which makes doctors hesitant to trust them.

The New Idea: Teaching the Detective the Rules of the Game

The authors of this paper created a new tool called EIGNN (Epidemiology-Informed Graph Neural Network).

Think of EIGNN as taking that super-smart detective and giving them a textbook on how viruses actually work. Instead of just guessing based on patterns, the detective is now forced to follow the "laws of physics" for diseases.

In the world of physics, if you drop a ball, gravity pulls it down. In the world of viruses, if a healthy person meets a sick person, there is a mathematical chance they will get sick. The authors built these mathematical rules (called ODEs or differential equations) directly into the detective's brain.

How It Works (The Analogy)

1. The Map (The Graph): The system looks at the hospital as a living map. Patients and rooms are dots, and the lines connecting them show who visited whom.
2. The Detective (The GNN): The AI looks at this map to understand the current situation.
3. The Rulebook (The Epidemiology): The AI is also given a rulebook that says things like: "A person cannot go from 'Healthy' to 'Recovered' instantly; they must be 'Sick' first."
4. The Training: The AI tries to predict who gets sick. If it makes a prediction that breaks the rules in the rulebook (like predicting someone recovered without being sick), the system scolds it and makes it try again. This forces the AI to learn both the patterns in the data and the biological rules.

What They Found

The researchers tested this new "Rule-Bound Detective" on three different scenarios:

1. Real Data: A real hospital during the COVID-19 pandemic.
2. Simulated Data: Computer-generated hospitals with viral and bacterial infections.

The Results:

• Better Accuracy: The new AI was better at predicting who would get sick, especially for longer timeframes (predicting 3 to 7 days ahead). It reached a success rate of nearly 98% in some tests.
• Trustworthy: Because the AI follows the biological rules, its predictions make more sense to doctors. It doesn't just say "this person is sick"; it explains the transition in a way that matches how diseases actually spread.
• Learning the Rules: The AI was so good that it could actually "learn" the rules of the disease on its own. For example, it figured out that the average recovery time was about 12 days, which matched the real-world data perfectly. It even calculated how contagious the virus was, matching known scientific values.

The "Continual Learning" Twist

The paper also found something interesting about how the AI learns over time.
Imagine if the detective only studied a map from last month and tried to predict next month. They would fail because the city changes.
The researchers found that if they let the AI continually update its knowledge as new data comes in (like a detective who updates their map every single day), it became incredibly robust. In fact, this daily updating was sometimes even more important than the rulebook itself, helping the AI handle the messy, changing reality of a real hospital.

The Bottom Line

This paper introduces a way to make AI for hospitals smarter and more trustworthy. By forcing the AI to respect the biological rules of how diseases spread, they created a tool that is not only accurate but also explainable. It's like giving a super-computer a biology degree so it can help doctors stop infections before they start.

Important Note: The paper focuses on predicting and interpreting these infections. While it suggests this could help with decisions like isolation or testing, the paper itself presents this as a tool for risk assessment and understanding disease dynamics, not as a finished clinical product ready for immediate use in every hospital.

Technical Summary

Technical Summary: Epidemiology-Informed Graph Neural Networks for Predicting and Interpreting Transmissible Hospital-Acquired Infections

Problem Statement
Transmissible hospital-acquired infections (HAIs) arise from complex, time-varying interactions among patients, healthcare workers, and clinical environments. While data-driven approaches like Graph Neural Networks (GNNs) effectively model these contact networks, they often function as "black boxes" that overlook established epidemiological principles. This lack of mechanistic constraints limits interpretability and clinical trust. Furthermore, existing epidemiology-informed learning methods primarily target population-level forecasting rather than patient-level infection risk prediction, leaving a gap in tools for individualized risk assessment and actionable infection prevention and control (IPC) strategies.

Methodology
The authors propose EIGNN (Epidemiology-Informed Graph Neural Network), a framework designed for patient-level state transition prediction in dynamic hospital settings. The approach integrates mechanistic epidemiological models into GNNs through a multi-task learning paradigm.

• Framework Architecture: The model processes a sequence of temporal graph snapshots representing the hospital contact network. A GNN node encoder generates per-individual embeddings from dynamic contact data. These embeddings are jointly used for four objectives:

  1. State Transition Prediction: Classifying individual state transitions (e.g., Stability, Progression, Resolution) over forecast horizons ($W \in \{1, 3, 7\}$ days).
  2. State Estimation: Predicting the specific epidemiological state (e.g., Susceptible, Infected, Recovered).
  3. Epidemiological Parameter Learning: Unsupervised estimation of key parameters (transmission rate $\beta$, recovery rate $\gamma$, etc.).
  4. Mechanistic Consistency: Enforcing adherence to governing differential equations.

• Loss Function: The training objective combines supervised losses for transitions ($L_{trans}$) and states ($L_{state}$) with unsupervised epidemiological constraints ($L_{epi}$) and transition constraints ($L_{const}$):
  $$L_f^t = \nu_{trans} L_{trans} + \nu_{state} L_{state} + \nu_{epi} L_{epi} + \nu_{const} L_{const}$$

• Epidemiology-Informed Strategies: Two distinct methods are implemented to embed mechanistic priors:

  1. AutoDiff: Enforces consistency with continuous-time Ordinary Differential Equations (ODEs) via automatic differentiation, analogous to Physics-Informed Neural Networks (PINNs).
  2. ODE-NSP (Next-State Prediction): Utilizes discrete-time approximations of ODEs (based on exponential waiting-time frameworks) as a differentiable physics layer to predict the next state, enforcing consistency with hazard equations.

• Constraints: The framework includes specific constraints to ensure biological plausibility:

  • Masked Loss ($L_{mask}$): Penalizes transitions that are biologically infeasible within a single time step (e.g., Susceptible $\to$ Recovered without passing through Infected).
  • Monotonicity Loss ($L_{mono}$): Enforces temporal consistency in disease progression scores.
  • Continual Learning (CL): Models are evaluated under a continual learning paradigm where parameters are updated sequentially as new data arrives, simulating non-stationary epidemic dynamics.

• Datasets: Evaluation was conducted on:

  • HUG-COVID: A real-world cohort of 282 hospitalized patients from Geneva University Hospitals (2020).
  • SocioPatterns: Pseudo-synthetic datasets derived from real contact networks (SFHH conference and Lyon hospital ward) with simulated SEIR dynamics.
  • Murcia: Fully synthetic datasets simulating Clostridioides difficile transmission in a hospital environment using a SEIRD-NS model.

Key Results

• Predictive Performance: On the real-world HUG-COVID dataset, the ODE-NSP strategy generally achieved the highest predictive performance, particularly at longer forecast horizons (3–7 days). For instance, at $W=7$ days, ODE-NSP with GraphSAGE achieved a balanced accuracy of 81.53% and an AUROC of 96.79%, outperforming non-informed baselines (75.01% and 92.84%, respectively).
• Horizon Dependency: The benefit of epidemiological constraints was most pronounced at intermediate-to-long horizons. At short horizons (1 day), strong non-informed backbones (e.g., GraphSAGE) performed well, leaving limited room for improvement.
• Continual Learning Impact: Continual learning (CL) significantly improved robustness, particularly for the AutoDiff method, which often suffered from instability under static training. CL mitigated AutoDiff's calibration issues, with some configurations achieving balanced accuracies above 92%.
• Parameter Recovery: The framework successfully learned epidemiological parameters in an unsupervised manner. On HUG-COVID, the model estimated a recovery period of ~12 days (consistent with the 14-day threshold) and a basic reproduction number ($R_0$) of 3.68. On synthetic datasets, the models demonstrated the ability to recover parameter directions and reproduce infection dynamics, though a trade-off was observed between exact parameter recovery and dynamic fidelity.
• Architecture Sensitivity: The efficacy of constraints varied by architecture. While ODE-NSP generally benefited GraphSAGE and TGN, it showed mixed results with STM-GNN, suggesting that constraints may conflict with internal memory mechanisms in complex architectures.

Significance and Claims
The paper claims that EIGNN offers a transparent tool for infection prevention and control by bridging the gap between data-driven prediction and mechanistic epidemiology. Key contributions include:

1. Patient-Level Focus: Unlike prior epidemiology-informed works focused on population-level forecasting, this framework targets individual risk assessment in dynamic hospital networks.
2. Interpretability and Trust: By embedding mechanistic priors (SIR, SEIR, SEIRD-NS), the model provides insights into pathogen dynamics and recovers meaningful epidemiological parameters directly from contact data, enhancing clinical interpretability.
3. Robustness: The integration of mechanistic constraints acts as a conditional inductive bias, stabilizing predictions in noisy, partially observable clinical environments, particularly when forecasting beyond immediate contact windows.
4. Actionable Insights: The framework supports risk-stratified decision-making (e.g., prioritizing testing or isolation) and enables prospective scenario analysis for IPC policies.

The authors maintain a modest stance, acknowledging limitations such as the reliance on one real-world dataset with limited patient-level features, the challenges of validating learned parameters in stochastic simulations, and the sensitivity of the framework to the validity of the underlying ODE formulations. They conclude that robust infection-risk prediction emerges from the interaction of expressive temporal graph architectures, mechanistic priors, and adaptive training strategies.