CTBN-PH: A continuous-time Bayesian network for individualised diagnostic risk prediction.

Area of Science:

• * Computational Biology
• * Medical Informatics
• * Biostatistics

Background:

• * Continuous-time Bayesian networks (CTBNs) offer dynamic modeling of disease progression but often lack patient-specific predictions.
• * Covariates significantly influence diagnostic transitions, posing a challenge for standard CTBN applications.
• * Integrating covariate effects into CTBNs is crucial for individualizing disease trajectory predictions.

Purpose of the Study:

• * To introduce the CTBN-PH model, combining CTBNs and Cox Proportional Hazards (Cox-PH) models.
• * To enable individualized prediction of complex disease trajectories by incorporating covariate effects.
• * To leverage causal topologies from healthcare data for dynamic risk estimation.

Main Methods:

• * Integration of CTBNs with Cox-PH models to form the CTBN-PH model.
• * Learning causal topologies from extensive healthcare trajectories (over 2.1 million patients).
• * Dynamic adjustment of transition intensities based on covariate effects for personalized risk assessment.

Main Results:

• * The CTBN-PH model successfully learned complex causal structures related to multi-morbid conditions like diabetes and hypertension.
• * Achieved an Integrated Brier Score (IBS) of 0.153 for predicting single diagnosis onset over 25 years.
• * Demonstrated strong performance in forecasting system inertia (IBS of 0.04 over four years) compared to non-individualized models.
• * Validated utility in simulating patient trajectories tailored to specific covariate-defined populations.

Conclusions:

• * The CTBN-PH model provides a robust framework for individualized disease trajectory prediction.
• * Incorporating causal inference and covariate effects significantly enhances predictive accuracy.
• * The model offers valuable applications in clinical decision support and personalized medicine simulations.