Model utility and explainability in federated learning - A case study in healthcare using fundus oculi datasets

Area of Science:

• Healthcare AI
• Machine Learning in Medicine
• Medical Imaging Analysis

Background:

• Patient privacy and data scarcity are significant hurdles in healthcare AI.
• Federated learning (FL) offers a solution by training models across decentralized datasets without sharing raw patient data.
• Evaluating FL in realistic, heterogeneous medical settings is crucial.

Purpose of the Study:

• To assess features learned by FL methods in a simulated, diverse healthcare setting with realistic data heterogeneity.
• To analyze the medical relevance of learned representations using explainability techniques.
• To evaluate FL performance against centrally trained models and identify robustness against common biases.

Main Methods:

• Combined six fundus oculi datasets to simulate a diverse FL environment.
• Evaluated three FL methods against centrally trained models for predictive performance and learned representations.
• Employed local and global explainability techniques, comparing local explanations to ophthalmologist-annotated attention maps.
• Assessed robustness against common biases in fundus datasets.

Main Results:

• Federated learning methods improved model utility by up to 9.97% compared to isolated training.
• Federated models learned the vertical cup-to-disc ratio, a key glaucoma diagnostic feature.
• High agreement was found between local explanations and ophthalmologist annotations, indicating medically relevant feature learning.
• Federated models demonstrated robustness against common dataset biases.

Conclusions:

• Federated learning systems offer significant benefits in healthcare, enhancing model utility while respecting patient privacy.
• This case study validates FL beyond idealized benchmarks, demonstrating its potential for learning medically relevant features.
• FL can mitigate spurious correlations and improve diagnostic accuracy in privacy-sensitive medical domains.