CALYPSO: Forecasting and Analyzing MRSA Infection Patterns with Community and Healthcare Transmission Dynamics

Rituparna Datta, Jiaming Cui, Gregory Madden, Anil VullikantiarXiv 2025 · DOI · arXiv

Introduction

Hospital-Acquired Infections (HAIs), particularly Methicillin-resistant Staphylococcus aureus (MRSA), represent a significant public health challenge, contributing to thousands of annual deaths and substantial healthcare expenditures. A critical complexity in managing MRSA is its transmission dynamic: it is not confined to clinical environments but spreads through a complex network of long-term care facilities (LTCFs) and the broader community via patient transfers and daily commuting.

Existing forecasting methods generally fall into two categories: mechanistic models (e.g., SIR models) and data-driven models (e.g., LSTMs). While mechanistic models offer interpretability and the ability to simulate "what-if" scenarios, they are difficult to calibrate with heterogeneous real-world data. Conversely, data-driven models excel at short-term forecasting but lack epidemiological grounding and the capacity for policy simulation. To bridge this gap, we present CALYPSO (CALibrated dYnamical Platform for Spatiotemporal Outbreaks), a hybrid framework that integrates the predictive strength of deep learning with the structural interpretability of metapopulation disease models.

The CALYPSO Framework

CALYPSO is designed as an end-to-end trainable system comprising three primary modules:

  1. Calibration Module ($NN_C$): An attention-based encoder-decoder network that processes historical time-series and static metadata to dynamically estimate disease transmission parameters, such as infection and recovery rates.
  2. Metapopulation Model ($M$): A compartmental SIRS (Susceptible-Infected-Recovered-Susceptible) simulator. It utilizes the parameters from $NN_C$ and a spatial "travel matrix"—derived from commuting and healthcare transfer data—to simulate MRSA spread across discrete "patches" (counties and facilities).
  3. Adapter ($NN_A$): A residual GRU network that fine-tunes the mechanistic model's outputs, correcting short-term errors without necessitating a full system retrain.
The CALYPSO architecture combining a calibration neural network, a mechanistic SIRS model, and a residual adapter.

The framework employs a multi-resolution loss function, optimizing for accuracy at the local patch, regional, and statewide levels simultaneously.

Data Integration

The study utilizes comprehensive datasets from the state of Virginia (2016–2021):

  • Virginia All-Payer Claims Database (APCD): Provides de-identified patient-level claims, including MRSA/MSSA diagnoses, demographics, and facility transfer records.
  • U.S. Census & ACS Data: Used for population normalization and building the mobility/travel matrix based on commuting patterns.
  • Virginia Department of Health Mapping: Facilitates the aggregation of counties into five distinct health regions.

Experimental Results

CALYPSO demonstrates superior performance over traditional baselines. At a 4-week forecast horizon, the framework improved statewide $R^2$ by approximately 4.5% and reduced Mean Squared Error (MSE) by 18% compared to a standard LSTM. When compared to a traditional mechanistic model calibrated via Ensemble Adjustment Kalman Filter (EAKF), CALYPSO achieved a reduction in MSE of over 95%.

Regional forecasts and calibrated simulation results across 644 patches in Virginia.

The model's ability to capture regional trends is evidenced by high $R^2$ values across various health regions, ranging from 0.82 in the Central region to 0.96 in the Southwest. Furthermore, the calibrated simulation results closely align with ground truth data across 644 distinct patches.

Policy Implications and Resource Allocation

Beyond forecasting, CALYPSO serves as a powerful tool for policy optimization. Key findings include:

  • Sensitivity to Data Quality: A mere 3% noise in healthcare facility data can significantly degrade statewide predictions. However, the model reveals that correcting data for a small number of high-impact facilities can restore most of the lost accuracy.
  • Intervention Spillover: Simulations show that reducing infection rates in one region can lead to localized increases in neighboring regions due to mobility-driven spillover, highlighting the necessity of coordinated, multi-facility strategies.
  • Efficient Resource Allocation: Using a greedy heuristic ("UnitGreedy"), the authors identified optimal facilities for intervention. Targeting just five facilities reduced simulated statewide infections by over 6,190 cases, a 2.7x improvement over targeting a single facility.

Conclusion and Future Work

While CALYPSO provides a robust framework for outbreak analysis, it is limited by the lack of exact recovery dates in insurance claims, necessitating literature-based assumptions. Future work involves the direct deployment of CALYPSO into state health agency pipelines to provide real-time forecasts and counterfactual simulations for resource allocation.

Summary updated

BibTeX
@misc{datta2025calypso,
  author = {Rituparna Datta and Jiaming Cui and Gregory Madden and Anil Vullikanti},
  title = {{CALYPSO: Forecasting and Analyzing MRSA Infection Patterns with Community and Healthcare Transmission Dynamics}},
  year = {2025},
  eprint = {2508.13548},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.13548},
}
Figure