[Paper Review] Learning the Probability of Activation in the Presence of Latent Spreaders

Maggie Makar, John V. Guttag+1 more|arXiv (Cornell University)|Dec 2, 2017

Data-Driven Disease Surveillance4 citations

TL;DR

This paper proposes PALS, a generative probabilistic model that learns the probability of infection by jointly modeling individual susceptibility and latent neighbor spreader states in networks. Using variational inference on patient contact networks, PALS outperforms benchmarks in predicting healthcare-associated C. diff infections and identifies high-risk, asymptomatic carriers—enabling targeted infection control interventions.

ABSTRACT

When an infection spreads in a community, an individual's probability of becoming infected depends on both her susceptibility and exposure to the contagion through contact with others. While one often has knowledge regarding an individual's susceptibility, in many cases, whether or not an individual's contacts are contagious is unknown. We study the problem of predicting if an individual will adopt a contagion in the presence of multiple modes of infection (exposure/susceptibility) and latent neighbor influence. We present a generative probabilistic model and a variational inference method to learn the parameters of our model. Through a series of experiments on synthetic data, we measure the ability of the proposed model to identify latent spreaders, and predict the risk of infection. Applied to a real dataset of 20,000 hospital patients, we demonstrate the utility of our model in predicting the onset of a healthcare associated infection using patient room-sharing and nurse-sharing networks. Our model outperforms existing benchmarks and provides actionable insights for the design and implementation of targeted interventions to curb the spread of infection.

Motivation & Objective

• To address the challenge of predicting infection risk when neighbor contagion states are latent and unobserved.
• To model both individual susceptibility (via observed characteristics) and exposure to hidden spreaders in a networked population.
• To improve risk estimation beyond existing models that ignore latent spreaders or fail to model individual-specific susceptibility.
• To identify actionable insights for infection control by learning which patient characteristics correlate with high transmission potential.
• To evaluate the model’s performance on synthetic data and real-world hospital patient data for C. diff transmission.

Proposed method

• Formulates a generative probabilistic model where infection is a function of individual susceptibility and latent Bernoulli-distributed spreader states of neighbors.
• Models each neighbor’s spreader state as a function of their observed characteristics (e.g., antibiotics, diarrhea) via learned weights u.
• Applies variational inference to jointly estimate latent spreader states and model parameters, enabling scalable inference.
• Uses a network structure where edges represent shared rooms or caregivers, encoding exposure pathways.
• Incorporates both susceptibility (e.g., age, medical history) and exposure (via neighbor influence) as inputs to predict infection probability.
• Employs a logistic regression-like link function to map characteristics and exposure to infection risk, with latent spreader states inferred from data.

Experimental results

Research questions

• RQ1 Can a model accurately predict infection risk when neighbor contagion states are latent and unobserved?
• RQ2 How does the inclusion of latent spreader states improve prediction accuracy compared to models that assume observed or ignored spreader status?
• RQ3 Which patient characteristics are most predictive of being a latent spreader of C. diff in a hospital setting?
• RQ4 To what extent can the model identify asymptomatic carriers who are not clinically diagnosed but still transmit the pathogen?
• RQ5 How does the model’s performance vary across different modes of infection (exposure-only, susceptibility-only, or combined)?

Key findings

• PALS achieved an AUC of 0.705 (95% CI: 0.703–0.706) on the nurse-sharing network with partial observation of spreader states, significantly outperforming benchmarks like NbrProbInf (AUC 0.641) and NbrInfRate (AUC 0.606).
• In the susceptibility-only setting, PALS achieved an AUC of 0.698, demonstrating robustness even when exposure plays no role.
• The model identified that patients on broad-spectrum antibiotics (e.g., IV-Vancomycin, Ciprofloxacin) and those treated for diarrhea (e.g., Loperamide) had the highest positive u weights, indicating elevated transmission potential.
• Patients receiving treatment for CDI had the largest negative u weights, suggesting that clinical isolation and contact precautions effectively reduce transmission risk.
• PALS achieved a true positive rate (TPR) of 0.324 at a false positive rate of 0.1 when using nurse-sharing data with partial observation, outperforming all benchmarks.
• The model accurately identified latent spreaders even when only partial information about neighbor states was available, with performance gains increasing under partial observability.