Transfer-learning on federated observational healthcare data for prediction models using Bayesian sparse logistic...

Area of Science:

• Biostatistics
• Machine Learning
• Clinical Prediction Modeling

Background:

• Small-sample clinical prediction problems often lack sufficient data for robust model development.
• Traditional methods struggle with data scarcity, leading to suboptimal model performance.
• Transfer-learning offers a promising avenue to improve models by utilizing external data.

Purpose of the Study:

• To develop a transfer-learning Bayesian sparse logistic regression model.
• To facilitate model fitting in small-sample clinical prediction tasks using informed priors.
• To transfer information from large datasets to small datasets effectively.

Main Methods:

• A Bayesian logistic regression framework incorporating transfer-learning was proposed.
• An informed, hierarchical prior was designed as a mixture of Bayesian Bridge shrinkage and normal distributions.
• Model performance was evaluated against traditional methods using metrics like AUC, calibration, bias, and sparsity.

Main Results:

• The transfer-learning model consistently outperformed traditional L1-regularized models in discrimination, calibration, bias, and sparsity.
• Even a continuous shrinkage prior alone improved performance over L1-regularization.
• The informed prior approach demonstrated superior results across simulations and a real-world clinical dataset.

Conclusions:

• Transfer-learning with informed priors effectively enhances prediction models in data-limited clinical settings.
• This method allows for privacy-preserving knowledge transfer as priors are independent of patient-level data.
• Future applications include cross-database learning and prediction for rare outcomes.