Federated Propensity Score Matching for Bias Correction in ICU.

Area of Science:

• Medical Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Multi-site medical studies face challenges in addressing confounding bias.
• Data privacy regulations restrict the centralization of sensitive patient information.
• Existing methods may not adequately balance bias reduction and data privacy.

Purpose of the Study:

• To propose a federated learning method integrating propensity score matching (PSM) for multi-site medical studies.
• To simultaneously address confounding bias and maintain data privacy.
• To evaluate the efficacy of within-hospital and cross-hospital PSM strategies within a federated learning framework.

Main Methods:

• Federated learning framework utilizing XGBoost algorithm.
• Integration of propensity score matching (PSM) for bias adjustment.
• Evaluation using data from five intensive care unit (ICU) databases (N=160,752).
• Comparison of within-hospital and cross-hospital PSM strategies.

Main Results:

• Cross-hospital PSM achieved 76.3% mean bias reduction (SMD: 0.070).
• Within-hospital PSM achieved 74.3% mean bias reduction (SMD: 0.074).
• Both strategies significantly reduced bias compared to baseline (SMD: 0.316).
• Predictive performance was maintained with AUROC values ranging from 0.73 to 0.75.

Conclusions:

• Federated learning with PSM is a viable method for causal inference in distributed healthcare networks.
• The proposed approach effectively reduces confounding bias without centralizing sensitive patient data.
• This method supports rigorous analysis in multi-site studies while adhering to privacy regulations.