Local and Distributed Machine Learning for Inter-hospital Data Utilization: An Application for TAVI Outcome...

Area of Science:

• Medical Prognostics
• Machine Learning
• Health Informatics

Background:

• Machine learning (ML) models are crucial for medical prognostics but often limited by single-center data.
• Data sharing across institutions is complex due to privacy and regulatory hurdles.
• Multi-center data integration enhances prognostic model robustness and accuracy.

Purpose of the Study:

• To develop accurate 1-year Transcatheter Aortic Valve Implantation (TAVI) mortality prediction models.
• To overcome data sharing impediments using distributed ML and local learning.
• To validate these techniques on data from two independent centers without sharing patient information.

Main Methods:

• A distributed ML framework with local learning and subsequent model integration was employed.
• Two patient cohorts (Center A: 1,160 patients, Center B: 631 patients) were used for TAVI mortality prediction.
• Five traditional ML algorithms were implemented and compared against mono-center models.

Main Results:

• Combined learning techniques demonstrated superior performance compared to mono-center models.
• Distributed XGBoost on Center A data improved AUC from 0.65 to 0.67.
• Distributed neural networks on Center B data improved AUC from 0.64 to 0.68.

Conclusions:

• Distributed ML and combined local models effectively address data sharing limitations in medical prognostics.
• This approach yields more accurate TAVI mortality estimation models.
• The methodology offers a viable solution for enhancing prognostic accuracy with limited multi-center data.