Distributed learning: Developing a predictive model based on data from multiple hospitals without data leaving the...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Oncology Data Science

Background:

• Personalized medicine requires extensive patient data for predictive modeling.
• Data sharing across hospitals faces ethical, legal, and administrative challenges.
• Distributed learning offers a solution by analyzing data locally.

Purpose of the Study:

• To demonstrate the feasibility of distributed learning for multi-institutional data analysis.
• To develop a predictive model for dyspnea in lung cancer patients using distributed learning.
• To overcome data sharing barriers in federated learning for healthcare.

Main Methods:

• A Bayesian network model was adapted for distributed learning.
• Data from 287 lung cancer patients across 5 medical institutes were utilized.
• The model was trained on data without it leaving individual hospitals.

Main Results:

• Successfully trained a Bayesian network model using distributed learning across multiple hospitals.
• Achieved an AUC of 0.61 (95% CI, 0.51-0.70) on cross-validation.
• External validation AUC ranged from 0.59 to 0.71, confirming model robustness.

Conclusions:

• Distributed learning facilitates multi-hospital model training while respecting data privacy.
• This approach enables knowledge extraction from routine patient data, complying with privacy laws.
• Distributed learning is a viable strategy for advancing personalized medicine and collaborative research.