A multicenter random forest model for effective prognosis prediction in collaborative clinical research network.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Privacy

Background:

• Prognostic prediction models are crucial for clinical decision-making.
• Individual institutions often have insufficient data for robust model development.
• Sharing sensitive biomedical data for multicenter studies presents significant privacy challenges.

Purpose of the Study:

• To develop a multicenter random forest (RF) prognosis prediction model using federated learning.
• To enable collaborative data mining from horizontally partitioned datasets while preserving patient privacy.
• To enhance RF model performance in multicenter settings compared to centrally trained models.

Main Methods:

• A novel differentially private generative adversarial network (GAN) was used for data enhancement.
• Federated learning enabled multicenter data mining without raw data aggregation.
• An importance ranking step was incorporated for privacy-preserving feature selection.

Main Results:

• The proposed model achieved performance comparable to or better than centrally trained RF models.
• The model demonstrated strong discrimination and calibration abilities on colorectal cancer datasets.
• Feature importance ranking provided insights without sharing patient-level data.

Conclusions:

• The federated RF model effectively overcomes performance limitations due to privacy constraints in multicenter clinical data analysis.
• The approach offers a practical solution for building reliable prognosis prediction models in collaborative research networks.
• This work addresses real-world challenges in medical artificial intelligence applications.