A dynamic model to estimate risk of major bleeding after PCI, trained and validated using data from the ACC's CathPCI Registry, exhibited less predictive error than static risk prediction models, according to a recent study published in PLOS Digital Health.

Nathan C. Hurley, MD, PhD, et al., included 2,868,808 PCIs from July 2009 to April 2015; 81% of cases were used to train the six tree-based machine learning models and the rest were used for validation. The models considered three key decision points in estimating bleeding risk: choice of access site, prescription of medication prior to PCI and choice of closure device. The primary outcome was in-hospital bleeding within 72 hours of the procedure.

When using presentation variables only, the area under the receiver operating curve (AUROC) was 0.812. Discrimination of the models improved to an AUROC of 0.845 when using all variables.

The initial model classified 123,712 patients as low bleeding risk. After incorporating all variables, 14,441 of these patients were reclassified as moderate risk and 1.4% of them experienced bleeds. An additional 723 patients were reclassified as high risk, 13% of whom experienced bleeds.

"As more data becomes progressively available, the model better identifies associations between variables and subsequent bleeding," write the authors. "The staged nature better represents an individual's risk throughout the course of treatment and can inform treatment decisions in a way that is superior to using a pre-PCI model in isolation."

Hurley and colleagues also highlight the opportunity for dynamic models like these to become integrated into electronic health records, adding, "While we segmented the data on key decision points within the registry data, we demonstrate the potential for implementing a clinical decision making tool within the medical record for real-time use, where each available data point can be added for key stages within each hospital system's clinical workflow."