Risk stratification in pulmonary arterial hypertension using Bayesian analysis

Area of Science:

• Cardiology
• Pulmonology
• Machine Learning
• Biostatistics

Background:

• Current risk stratification tools for pulmonary arterial hypertension (PAH) have limited predictive accuracy.
• This limitation stems from the assumption of independent and linear relationships between prognostic variables and outcomes.
• There is a need for enhanced risk prediction models in PAH management.

Purpose of the Study:

• To demonstrate the utility of Bayesian network-based machine learning in improving PAH risk stratification.
• To enhance the predictive ability of the existing REVEAL 2.0 risk stratification tool.
• To develop a novel risk prediction model named PHORA.

Main Methods:

• A tree-augmented naïve Bayes model (PHORA) was developed using data from the REVEAL registry.
• PHORA utilizes the same variables and cut-points as the REVEAL 2.0 tool for predicting 1-year survival.
• Internal and external validation was performed across the REVEAL, COMPERA, and PHSANZ registries.

Main Results:

• PHORA achieved an Area Under the Curve (AUC) of 0.80 for 1-year survival prediction, outperforming REVEAL 2.0 (AUC 0.76).
• External validation showed AUCs of 0.74 (COMPERA) and 0.80 (PHSANZ).
• PHORA demonstrated excellent discrimination between low-, intermediate-, and high-risk groups in all registries.

Conclusions:

• The Bayesian network-derived PHORA model shows improved risk prediction discrimination in PAH.
• Bayesian networks effectively account for interrelationships between clinical variables and outcomes.
• PHORA offers enhanced predictive capabilities and tolerance to missing data in PAH risk assessment.