Dynamic prediction modeling approaches for cardiac surgery.

Area of Science:

• Cardiovascular Surgery
• Health Services Research
• Biostatistics

Background:

• Cardiac prediction models often experience calibration deterioration over time.
• Accurate risk assessment for in-hospital mortality post-cardiac surgery is crucial.
• Patient populations are heterogeneous, necessitating adaptable prediction tools.

Purpose of the Study:

• To compare various model fitting approaches for predicting in-hospital mortality after cardiac surgery.
• To assess methods that adjust for evolving case mix in a diverse patient population.
• To evaluate strategies for maintaining model calibration over time.

Main Methods:

• Analysis of over 300,000 cardiac surgery procedures from England and Wales (2001-2011).
• Comparison of model fitting approaches: no update, periodic refitting, rolling window, and dynamic logistic regression.
• Covariate adjustment using variables from the logistic European System for Cardiac Operative Risk Evaluation (EuroSCORE) model.

Main Results:

• The association between in-hospital mortality and certain variables has changed over time.
• A decreasing intercept coefficient indicates reduced observed mortality.
• Risk factor associations varied, with some stable (e.g., operative urgency) and others changing (e.g., left ventricular function).

Conclusions:

• Dynamic models or periodic refitting are necessary to address calibration drift.
• Dynamic modeling frameworks offer continuous updates and allow for individual risk factor inference.
• Improved, time-resilient models enhance governance, quality improvement, and patient-level decision-making.