A Bayesian latent-class model framework to estimate disease burden of respiratory syncytial virus using imperfect and heterogeneous laboratory diagnostic data

This study proposes a novel Bayesian latent-class model that effectively estimates the disease burden of respiratory syncytial virus (RSV) by integrating heterogeneous, imperfect diagnostic data, demonstrating superior accuracy compared to existing methods when applied to large sample sizes to inform national immunization policies.

The Gist

The Big Problem: The "Invisible" Virus

Imagine you are trying to count how many people in a city have caught a specific cold virus (RSV). You send out a team of detectives (the lab tests) to check people who are sick.

But here's the catch: The detectives aren't perfect.

1. They miss things: Sometimes a detective looks at a suspect and says, "You're clean," even though they actually have the virus. This happens because the virus might be hiding deep in the lungs, or the detective arrived too late after the person got sick.
2. They use different tools: Some detectives use a high-tech scanner (a very sensitive test), while others use a basic flashlight (a less sensitive test).
3. They show up at different times: Some detectives arrive right when the suspect is most visible; others arrive days later when the suspect has already gone into hiding.

If you just count the people the detectives found, you will severely underestimate the total number of sick people. It's like trying to count fish in a lake by only looking at the ones that jump out of the water; you'll miss the thousands swimming deep below.

The Old Ways of Guessing

Before this paper, scientists tried to fix this "underestimation" problem in two ways:

1. The "Naïve" Count: Just counting the positive tests.
   • Analogy: This is like counting only the fish you see jumping. You get a number, but you know it's way too low.
2. The "Multiplier" Method: Taking the low number and multiplying it by a fixed factor (e.g., "Let's just multiply our count by 2 to guess the real number").
   • Analogy: This is like saying, "I saw 10 fish, so I'll guess there are 20." It's better than nothing, but it's a blunt instrument. It doesn't account for why the detectives missed the fish (was it the tool? was it the timing?). It assumes the same mistake happens every single time, which isn't true.

The New Solution: The "Super-Detective" (Bayesian Model)

The authors of this paper built a new, sophisticated computer model. Think of this model as a Super-Detective AI that doesn't just count the fish; it understands the behavior of the detectives and the fish.

Here is how it works, using our analogy:

• It knows the tools: The AI knows that the "high-tech scanner" finds 80% of the fish, while the "flashlight" only finds 50%.
• It knows the timing: The AI knows that if a detective arrives 5 days after the fish jumped, they are only 50% likely to see it.
• It looks at the whole picture: Instead of just looking at one person, it looks at the patterns of thousands of people. It asks: "If we have 30,000 detectives looking with these specific tools at these specific times, and we only found 500 fish, how many fish must actually be in the lake?"

The Big Discovery: Size Matters

The most important finding of this paper is about how much data you need to make the Super-Detective work.

• Small Data (Too few detectives): If you only have a small number of tests (like 2,500 or 5,000), the AI gets confused. It can't tell the difference between "there are very few fish" and "our detectives are just really bad at their job." In this scenario, the model might actually guess the number is too high (over-estimation) because it tries so hard to compensate for the missing data.
• Big Data (Many detectives): Once you have a lot of data (around 30,000 tests or more), the AI becomes incredibly accurate. It can finally separate the "bad detective" factor from the "low fish count" factor.
  • The Result: At 30,000 tests, the model is 80% accurate. At 60,000 tests, it's 95% accurate.

Why This Matters for Real Life

Why do we care about counting RSV in adults?

• Vaccines: We now have new vaccines for older adults. To decide who should get them (everyone over 75? only those with heart disease?), health officials need to know exactly how dangerous the virus is.
• Policy: If we use the old "Naïve" or "Multiplier" methods, we might think the virus isn't a big deal and skip vaccinating people who really need it.
• The Fix: This new model gives health officials a reliable way to say, "Okay, based on the messy, imperfect data we have, the real number of sick people is X." This helps them make better decisions to save lives.

The Bottom Line

The paper is essentially saying: "We built a smarter calculator to count the virus. It's much better than the old ways, but it needs a lot of data (at least 30,000 tests) to work its magic. If you give it enough data, it can see the invisible virus and help us protect the vulnerable."

Technical Summary

1. Problem Statement

Accurate estimation of the disease burden of Respiratory Syncytial Virus (RSV) in adults is significantly hindered by the limitations of current diagnostic practices. Key challenges include:

• Imperfect Testing Performance: Diagnostic tests have less than 100% sensitivity, leading to false negatives. Sensitivity varies based on the type of clinical specimen (e.g., nasopharyngeal swab, saliva, sputum, serum), the diagnostic test used, and the timing of specimen collection relative to symptom onset (viral load declines rapidly after 5 days).
• Heterogeneous Data: Real-world data often involve individuals receiving varying numbers of tests (1 to 4 different approaches) at different times.
• Limitations of Existing Methods:
  • Naïve Approach: Simply counts positive tests, ignoring false negatives and underestimating burden.
  • Multiplier Approach: Applies population-level upscaling factors. These often fail to account for individual-level heterogeneity in testing patterns and specimen timing, and they rely on a "practical" gold standard that is not truly 100% sensitive, leading to systematic underestimation.
  • Specificity Issues: Existing methods often ignore false positives (specificity < 100%), potentially overestimating burden in low-prevalence seasons, though the primary issue identified is underestimation due to sensitivity.

2. Methodology

The authors proposed a novel Bayesian latent-class model designed to integrate complex, individual-level diagnostic data.

• Model Structure:

  • Latent Class: The true RSV infection status of an individual is treated as a latent variable.
  • Covariates: The probability of infection is modeled as a function of age (four groups: 18–49, 50–64, 65–74, ≥75) and calendar month (to capture seasonality).
  • Testing Mechanism: The model accounts for up to four different testing approaches per individual. It incorporates:
    • Peak Sensitivity: An unknown parameter inferred for each test type.
    • Relative Sensitivity: A time-varying factor based on days since admission (e.g., 100% at 0–2 days, dropping to 50% at 6–14 days).
    • Specificity: Fixed at 0.998 based on literature.
  • Inference: Parameters were estimated using Hamiltonian Monte Carlo (HMC) via the Stan software (rstan package) with non-informative Beta priors.

• Simulation Framework:

  • Data Generation: Simulated datasets were created to mimic real-world RSV epidemiology in the UK across three respiratory seasons (Aug 2021–July 2024).
  • Scenarios: The study varied sample sizes (2,500 to 30,000 patients) and testing practices (e.g., test frequency proportional to sensitivity vs. equal distribution).
  • Comparison: The Bayesian model was benchmarked against the Naïve model and the Multiplier model.
  • Validation: Performance was assessed using calibration slopes (comparing estimates to known "true" simulated values) and the proportion of 95% Credible Intervals (CrIs) containing the true value.

3. Key Contributions

• Novel Framework: Developed the first modeling framework that simultaneously accounts for heterogeneous testing approaches, variable specimen collection times, and imperfect sensitivity/specificity at the individual patient level.
• Quantification of Bias: Systematically demonstrated that conventional methods (Naïve and Multiplier) systematically underestimate RSV burden regardless of sample size.
• Sample Size Guidelines: Established critical thresholds for study design, identifying that robust estimation requires significantly larger sample sizes than previously assumed for latent class models.
• Open Science: Provided the simulation code and data via a public GitHub repository to facilitate reproducibility and application in other contexts.

4. Key Results

• Sample Size Dependency:
  • Small Samples (<15,000 tests): The Bayesian model exhibited severe over-estimation (bias) because it could not reliably distinguish between low prevalence and low test sensitivity in sparse data.
  • Moderate Samples (30,000 tests): Accuracy improved significantly, with >80% of estimates falling within the true value range.
  • Large Samples (60,000 tests): The model achieved >95% accuracy.
• Comparison with Existing Models:
  • Naïve Model: Consistently underestimated burden (estimates were 38%–88% of true values).
  • Multiplier Model: Showed moderate improvement but still systematically underestimated burden across all sample sizes.
  • Bayesian Model: Outperformed both alternatives when sample sizes were sufficient (≥30,000 tests). Its advantage was even more pronounced when the number of available testing approaches was reduced (e.g., from 4 to 3).
• Calibration: As sample size increased, the calibration slope for peak sensitivity and RSV positive proportions approached 1.0 (perfect calibration), whereas small samples resulted in slopes indicating severe over-estimation.

5. Significance and Implications

• Public Health Policy: The study provides a robust tool for generating accurate RSV disease burden estimates, which is critical for informing national immunization policies, particularly regarding the expansion of RSV vaccine eligibility to broader adult populations (e.g., lowering age thresholds or including those with comorbidities).
• Study Design Guidance: The findings offer a practical guideline for researchers: to use Bayesian latent class models for RSV burden estimation, studies should aim for a minimum of 15,000 tests (approx. 7,500 patients), with 30,000 tests (15,000 patients) being the ideal target for robust, unbiased results.
• Methodological Advancement: By moving from population-level multipliers to individual-level Bayesian inference, the study sets a new standard for handling imperfect diagnostic data in epidemiology, applicable to other respiratory pathogens with similar testing challenges.

Conclusion: The proposed Bayesian latent-class model is a superior framework for estimating RSV burden in adults when applied to sufficiently large, heterogeneous datasets. It effectively disentangles true infection prevalence from measurement error, correcting the systematic underestimation inherent in current standard practices.