Machine Learning for Predicting Mortality in Transcatheter Aortic Valve Implantation: An Inter-Center Cross...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Current prognostic scores for transcatheter aortic valve implantation (TAVI) lack machine learning integration.
• Strict data sharing regulations hinder the application of advanced machine learning in healthcare.
• Accurate risk stratification for one-year mortality before TAVI is crucial.

Purpose of the Study:

• To develop and validate machine learning models for predicting one-year mortality in TAVI patients.
• To establish a robust, data-privacy-preserving model validation approach across different clinical centers.
• To assess the performance of machine learning models validated externally without direct patient data exchange.

Main Methods:

• A machine learning model was developed using data from one center.
• The model was externally validated at a second center using a novel exchange protocol that shared only models and processing pipelines, not patient data.
• Model robustness was evaluated on the external center's data, involving 1300 and 631 patients per center.

Main Results:

• Models developed with larger datasets demonstrated comparable or superior prediction accuracy on external validation data.
• Logistic regression, random forest, and CatBoost models achieved Areas Under the ROC Curve (AUC) of 0.65, 0.67, and 0.65 internally, and 0.62, 0.66, and 0.68 externally, respectively.
• The proposed exchange protocol effectively validated model performance across centers.

Conclusions:

• A scalable, privacy-preserving protocol enables robust external validation of machine learning models for TAVI risk stratification.
• This approach overcomes data sharing barriers, facilitating broader adoption of machine learning in clinical settings.
• The methodology can be extended to other clinical applications requiring multi-center model validation.