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

Background: Accurate estimation of respiratory syncytial virus (RSV) disease burden is challenged by the imperfect testing performance that varies by clinical specimens, diagnostic tests, and timing of specimen collection. Although the use of multiple testing approaches (such as testing multiple clinical specimens or additional diagnostic tests) could increase the RSV detection, there is absence of a modelling framework to fully incorporate the complexity of heterogeneous diagnostic data. In this study, we proposed a novel Bayesian latent class model that accounted for heterogeneous data on the number of RSV tests and variable specimen collection time among individual patients, imperfect testing sensitivity and specificity of different combinations of clinical specimen and diagnostic test (i.e., testing approaches), and RSV seasonality. Methods: Using simulated datasets consisting of four different testing approaches that mimic real-world RSV epidemiologic characteristics in the UK under different sample size and testing practice scenarios, we assessed the model performance in estimating RSV disease burden as the annual RSV positive proportion in lower respiratory tract infection (LRTI) cases across three respiratory seasons (August 2021 to July 2024) in four adult age groups: 18 to 49 years, 50 to 64 years, 65 to 74 years and over 75 years. Results: We demonstrated that model performance increased substantially with increased sample size, achieving over 80% in accuracy at a sample size of 30,000 tests and 95% in accuracy at a sample size of 60,000 tests; by contrast, smaller sample size could lead to severe over-estimation of the RSV disease burden. In comparison with the existing approaches, both the naive model and the multiplier model systematically under-estimated the RSV disease burden regardless of sample size. The Bayesian model yielded more accurate estimates when the sample size reached 30,000 tests or more; its advantage over the other two models was even more pronounced if the number of testing approaches reduced to 3. Conclusion: The findings above suggest that the proposed Bayesian model provides a robust framework for estimating RSV burden by integrating complex, individual-level testing data when fitting with sufficient input data, offering a critical tool for generating more accurate RSV disease burden estimates to inform national immunisation policies.