In a recent study published in the EClinicalMedicine, a group of researchers developed and validated a Chronic Obstructive Pulmonary Disease (COPD) (a progressive lung condition that causes breathing difficulties)-specific mortality risk prediction model using probabilistic graphical modeling to enhance disease management strategies.

Study: Development and validation of a mortality risk prediction model for chronic obstructive pulmonary disease: a cross-sectional study using probabilistic graphical modelling. Image Credit: Jo Panuwat D/Shutterstock.com

Background

COPD is a major global cause of mortality. Predictive models like Body mass index, Obstruction, Dyspnea, Exercise capacity (BODE), Age, Dyspnea, Obstruction (ADO), and Dyspnea, Obstruction, Smoking status, Exacerbation frequency (DOSE) help identify high-risk COPD patients, but these primarily focus on all-cause mortality.

Traditional models, such as regression and random survival forests, are limited to associative predictions lacking causal insight. In contrast, probabilistic graphs, or causal graphs, can identify potential cause-effect relationships from observational data by factoring out confounders.

Further research is needed to refine and validate COPD-specific mortality predictors across diverse populations and to explore underlying biological mechanisms for targeted interventions.

About the study 

The present study followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines, with all participants providing informed consent.

The discovery cohorts were drawn from the COPD Genetic Epidemiology (COPDGene) Study, which included 10,198 current and former smokers aged 45 to 80 years with over ten pack-years of smoking history.

Data collected included demographic, spirometry, clinical, and chest Computed Tomography (CT) scan features, along with all-cause and COPD-specific mortality, defined by criteria excluding deaths from comorbidities like cardiovascular disease (CVD) or cancer. The final analysis focused on 8,610 participants with complete baseline and follow-up data.

For external validation, the Evaluation of COPD Longitudinally to Identify Predictive Surrogate End-points (ECLIPSE) study, which recruited 2,501 participants from the United States (US) and Europe, was used, selecting 2,312 individuals with complete three-year mortality data.

Directed probabilistic graph models were constructed using the CausalCoxMGM method to identify direct predictors of mortality while accounting for confounders.

These models were compared to ADO, Updated BODE indices, and standard machine learning approaches, with performance evaluated via cross-validation using Harrell’s concordance index.

The Vital capacity-forced vital capacity (FVC) %predicted, Age, history of Pneumonia, Oxygen saturation, forced expiratory volume in 1s (FEV₁)/FVC Ratio, 6-min walk Exercise capacity, Dyspnoea (VAPORED) risk score was developed using seven features linked to COPD-specific mortality, and its accuracy was validated in the ECLIPSE cohort for predicting survival over one, two, and three years.

Study results 

In the Phase 2 study, which included a subset of Phase 1 participants, notable changes were observed in clinical covariates. These changes included an expected five-year increase in age and a significant reduction in patients in the more severe Global Initiative for Chronic Obstructive Lung Disease (GOLD) categories.

Additionally, there was a significant increase in the incidence of comorbidities, such as CVD and diabetes. The BODE index showed a significant decrease, while the ADO index increased in Phase 2 participants compared to Phase 1. Despite these changes, the survival functions between the two phases were not statistically different, indicating consistent all-cause mortality across both phases.

The ECLIPSE study, used for external validation, differed significantly from COPDGene, with a higher proportion of male participants and less racial diversity. The ECLIPSE cohort also included more severe COPD cases, as reflected in higher ADO, BODE, and updated BODE indices and a higher rate of all-cause mortality. This difference was significant, especially in the number of deaths observed within the first three years.

The study’s analysis identified features directly linked to COPD-specific mortality. While there was significant overlap in the variables affecting all-cause and COPD-specific mortality, some differences emerged.

For example, FVC %predicted was strongly associated with COPD-specific mortality, while FEV₁ %predicted was more relevant to all-cause mortality. Comorbidities like CVD and diabetes were linked only to all-cause mortality.

Graph-based predictive models, developed from these insights, outperformed traditional indices like ADO and updated BODE in predicting both all-cause and COPD-specific mortality. The models demonstrated the ability to stratify patients into distinct risk groups more effectively than the BODE index.

For external validation, the VAPORED risk score, developed using seven clinical variables, was tested on the ECLIPSE cohort. The VAPORED model significantly outperformed ADO and BODE, and updated BODE indices across several predictive metrics, particularly in the concordance probability estimate.

The model’s predictions were well-calibrated for one-, two-, and three-year survival probabilities in the ECLIPSE study.

Additionally, a web-based tool was developed to allow users to calculate and compare mortality risk using the VAPORED score, BODE, and ADO indices. This tool is accessible as a Shiny app, enabling clinicians and researchers to evaluate and visualize mortality risks based on key clinical variables.

Conclusions 

To summarize, this study used probabilistic graph modeling to identify features directly linked to COPD-specific and all-cause mortality, distinguishing them from simple correlates. The researchers developed the VAPORED risk score using clinical data, which outperformed traditional indices like ADO and BODE in predicting mortality.

Additionally, the study identified unique factors such as Internet access and specific biological markers associated with increased all-cause mortality risk. This approach demonstrated superior predictive power, suggesting new avenues for targeted interventions and the potential for developing more comprehensive mortality risk scores for COPD patients.