Abstract
Methicillin-resistant Staphylococcus aureus (MRSA) colonization across multiple body sites is a significant risk factor for infection. While decolonization protocols exist, their multi-component nature makes it difficult to determine which specific treatments drive success. We developed a Bayesian coupled hidden Markov model (CHMM) to analyze longitudinal data from the CLEAR trial. Our model identifies the nares as a central hub for colonization and quantifies the impact of individual therapies. We found that while all components contribute, nasal mupirocin is the most critical, and the protocol effectively reduces self-inoculation between sites. Simulations suggest that enhancing MRSA clearance at the skin would yield the largest gains to the overall decolonization regimen.
Introduction: The "Black Box" of Decolonization
MRSA is a resilient pathogen that inhabits various reservoirs, including the nose, throat, skin, and wounds. To mitigate infection risks, clinicians deploy "bundles" of treatments—such as nasal ointments and antiseptic washes. However, standard Randomized Controlled Trials (RCTs) typically evaluate the bundle as a whole, leaving the individual contribution of each component unknown. This "black box" approach makes it challenging to optimize protocols or identify redundancies. This research aims to bridge that gap by using machine learning to simulate the hidden dynamics of how MRSA moves between body sites during treatment.
Methods: Decoupling Site Dynamics
To look inside this black box, the authors formulated a Bayesian Coupled Hidden Markov Model (CHMM). Unlike standard HMMs that track a single sequence of states, a CHMM allows multiple sequences—representing different body sites—to interact over time. In this framework, the probability of a site becoming colonized depends not only on its own previous state but also on the colonization status of other sites in the previous time step.
The model was applied to longitudinal data from the CLEAR (Changing Lives by Eradicating Antibiotic Resistance) Trial, which followed 2,121 participants over six months post-hospital discharge. The dataset included over 20,500 swabs from the nares, skin, throat, and wounds. By using Markov Chain Monte Carlo (MCMC) sampling, the team estimated the "strength" of interactions between these sites and the specific "clearance" effect of each treatment component.
Key Insights: The Nares as a Colonization Hub
The model's most striking finding is the central role of the nares. In the control (education) group, colonization in the nose was consistently linked to the subsequent colonization of all other body sites, confirming its status as a primary reservoir or "hub."
The decolonization protocol achieves success through two primary mechanisms:
- Direct Clearance: It significantly increases the rate at which bacteria are eliminated from specific sites.
- Interruption of Self-Inoculation: It weakens the links between sites, preventing the "ping-pong" effect where MRSA moves from one reservoir (like the nose) to another (like a wound or skin).
One of the most practical findings was the discovery that chlorhexidine (CHG) mouthwash was the least effective component. The model estimated that even without the mouthwash, the protocol would retain 97% of its effectiveness, suggesting clinical protocols could be simplified without significantly losing efficacy.
Results and Optimization
The CHMM accurately mirrored the clinical trial results, predicting the observed drop in MRSA carriage from 66% to 48% in the education group and down to 26% in the decolonization group. Nasal mupirocin emerged as the single most powerful tool, independently capable of reducing carriage from 66% to 31%.
Furthermore, in silico experiments assessed the impact of hypothetical improvements. While the model confirmed the nares as the primary engine for persistence, sensitivity analyses suggested that enhancing MRSA clearance at the skin would yield the largest gains to the overall decolonization regimen, highlighting the importance of comprehensive body-site targeting.
Conclusion
By moving beyond simple outcome tracking to modeling the underlying dynamics of colonization, this study provides a roadmap for evidence-based clinical protocols. The ability to simulate "leaner" treatments offers a path toward higher patient adherence and more efficient healthcare delivery. The authors have released an R-package (chmmMCMC) to allow other researchers to apply these models to different infectious diseases, potentially streamlining protocols for a variety of pathogens.
Summary updated