TL;DR
This study applies Bayesian spatial survival models to understand factors influencing dengue hospitalization length of stay in Makassar, Indonesia. The research found that the Leroux prior generally outperformed other spatial priors, and while both Weibull and Cox models performed similarly, the Weibull model showed a slightly better predictive fit. Key patient factors like age and platelet count were associated with shorter hospital stays, while red blood cell count was linked to longer stays.
What Problem Does This Paper Address?
Understanding the duration of hospital stays for infectious diseases like dengue is crucial for effective hospital management, resource allocation, and public health interventions. While survival analysis is a standard tool for time-to-event data, incorporating spatial dependencies can provide a more nuanced understanding, especially when disease incidence or severity might vary geographically. Previous research on dengue hospitalization duration has often overlooked spatial correlations or relied on less sophisticated modeling techniques. This paper aims to fill this gap by applying and comparing advanced Bayesian spatial survival models to a real-world dengue dataset, specifically focusing on identifying influential factors and evaluating different spatial modeling approaches.
Methods
The researchers employed two primary modeling frameworks: a parametric Weibull survival model and a semiparametric Cox proportional hazards model. Both models were extended to incorporate spatial frailty terms, allowing for unobserved heterogeneity across different geographical areas (districts in this case). They specifically investigated three spatial prior formulations for these frailty terms:
- Leroux prior: This prior allows for a flexible adjustment of the strength of spatial autocorrelation between areas.
- Intrinsic Conditional Autoregressive (ICAR) prior: A common choice for spatial modeling, assuming a dependency structure based on adjacency.
- Independent prior: This serves as a baseline, assuming no spatial structure and essentially modeling an unstructured frailty.
The models were implemented using Bayesian inference with MCMC (Markov Chain Monte Carlo) methods via the R2WinBUGS package. Model comparison was performed using the Deviance Information Criterion (DIC) and Watanabe-Akaike Information Criterion (WAIC) for predictive fit. The significance of covariates was assessed by examining their 95% posterior credible intervals.
Datasets
The study utilized a dataset of inpatient dengue fever and dengue hemorrhagic fever cases admitted to Dr. Wahidin Sudirohusodo hospital in Makassar, Indonesia, between January 1, 2013, and July 31, 2018. The dataset included 705 patients after exclusions. Information collected comprised:
- Length of stay (LOS) in days.
- Patient discharge status (recovered, not recovered, died, transferred).
- Demographic information: sex, age.
- Clinical variables: white blood cell (WBC) count, red blood cell (RBC) count, hemoglobin (HGB) levels, hematocrit (HCT) percentage, and platelet (PLT) count.
- Residential addresses, geocoded to the district level within Makassar city.
Key Insights and Ideas
- The Leroux prior is a flexible and powerful tool for spatial survival analysis: It demonstrated superior performance compared to ICAR and independent priors, suggesting that modeling the strength of spatial autocorrelation is more beneficial than assuming fixed adjacency-based structures or no spatial structure at all.
- Weibull and Cox models offer similar insights but differ in fit: While both models identified the same key patient factors influencing LOS, the Weibull model provided a slightly better predictive fit according to DIC and WAIC. This highlights the value of considering parametric models when the underlying hazard function can be reasonably assumed.
- Spatial patterns in hospitalization duration might be subtle after covariate adjustment: Despite using spatial models, the study found that after accounting for key patient characteristics, the variation in LOS across districts was minimal. This suggests that individual patient factors play a more dominant role than geographical location in determining hospitalization duration in this specific context.
- Clinical and demographic factors are significant predictors of dengue hospitalization duration: Age and platelet count were negatively associated with LOS (shorter stays), while RBC count was positively associated (longer stays). This aligns with existing knowledge about dengue severity and its impact on patient recovery.
Key Results
- Model Performance: The Leroux prior was preferred for both Weibull and Cox models over ICAR and independent priors, as indicated by lower DIC and WAIC values. The Weibull model generally showed a slightly better predictive fit than the Cox model.
- Covariate Associations:
- Negative association with LOS (shorter stays): Age and Platelet (PLT) count.
- Positive association with LOS (longer stays): Red Blood Cell (RBC) count.
- These associations were consistent across all three spatial priors and both model types.
- Spatial Frailty: For most districts, the 95% credible intervals for spatial random effects (hazard ratios) included one, suggesting little significant spatial variation in LOS after covariate adjustment. This was particularly true for the Cox-Leroux and Weibull-Leroux models.
Limitations and Future Work
- Single Hospital Data: The study relied on data from a single hospital (Dr. Wahidin Sudirohusodo). The authors acknowledge that including data from other hospitals could reveal stronger spatial patterns and potentially different hospital-specific effects.
- Limited Spatial Resolution: The spatial analysis was conducted at the district level. Finer spatial resolutions might uncover more localized variations.
- Not All Dengue Cases Included: Hospitalization data only captures severe cases requiring admission, not all dengue infections. This means the findings might not represent the entire dengue burden in the population.
- Missing Data: A small percentage (6.8%) of clinical records had missing data, which could potentially influence the results.
Future work suggested includes:
- Expanding the analysis to include data from multiple hospitals to investigate inter-hospital variations and more robust spatial patterns.
- Comparing the duration of stay across different hospitals.
- Investigating the impact of other potential spatial factors or finer spatial resolutions.
Summary updated