Propensity Weighted federated learning for treatment effect estimation in distributed imbalanced environments.

Area of Science:

• Medical Informatics
• Causal Inference
• Machine Learning

Background:

• Estimating treatment effects from observational data is crucial in medicine due to data availability and the limitations of randomized controlled trials.
• Challenges arise in distributed causal inference, particularly when patient data cannot be shared between hospitals with differing patient distributions and treatment assignment criteria.
• Existing federated learning approaches may struggle with significant shifts in treatment distributions across decentralized nodes.

Purpose of the Study:

• To address the challenge of distributed causal inference in healthcare settings with privacy constraints and heterogeneous treatment distributions.
• To propose a novel federated learning algorithm adapted for estimating individual treatment effects in decentralized environments.
• To evaluate the performance of the proposed algorithm against existing methods under varying degrees of treatment distribution imbalance.

Main Methods:

• Adaptation of the FederatedAveraging algorithm to the Treatment Effect DEcision VAult Estimation (TEDVAE) neural network model.
• Development of Propensity Weighted Federated Averaging (PW FedAvg) to account for shifts in treatment assignment distributions across hospitals.
• Experimental comparison of PW FedAvg with Vanilla Federated Averaging and other federated learning causal inference algorithms.

Main Results:

• PW FedAvg demonstrates improved accuracy in estimating individual causal effects compared to baseline methods, especially when treatment distributions are highly imbalanced across hospitals.
• The proposed method effectively handles variations in patient populations and treatment assignment criteria without direct data sharing.
• Experimental results show significant error reduction in causal effect estimation with PW FedAvg under challenging imbalance scenarios.

Conclusions:

• PW FedAvg offers a robust solution for privacy-preserving distributed causal inference in healthcare.
• The algorithm's ability to manage treatment distribution shifts makes it suitable for real-world clinical settings with decentralized data.
• This work advances the application of federated learning for reliable treatment effect estimation in complex observational data scenarios.