Confederated learning in healthcare: Training machine learning models using disconnected data separated by individual, data type and identity for Large-Scale health system Intelligence

Dianbo Liu, Kathe P. Fox, Griffin M. Weber, Timothy A. MillerJ. Biomed. Informatics 2022 · OpenReview

TL;DR

This paper introduces "confederated machine learning," a novel approach to train AI models on fragmented healthcare data that is separated not only by patient but also by data type and patient identity, without requiring direct data sharing or patient ID matching. This method offers a practical solution for building large-scale health intelligence systems by overcoming significant privacy and logistical barriers in healthcare data utilization.

What Problem Does This Paper Address?

Healthcare data is notoriously fragmented, residing in isolated "silos" across different providers and institutions. This fragmentation occurs in multiple ways:

  • Horizontal Separation: Data for the same individual is spread across different healthcare systems or clinics.
  • Vertical Separation: Different types of data for the same individual (e.g., diagnoses, medication records, lab results) are stored separately.
  • Identity Separation: Even when data types are available, matching patient records across these silos is often impossible due to privacy concerns and inconsistent identification systems.

Traditional federated learning addresses horizontal separation by training models locally and aggregating updates. However, it struggles with vertical separation and identity separation. Existing methods for vertically separated data often require patient ID matching, which is a major hurdle in healthcare. This paper tackles the challenge of training machine learning models on data that suffers from all three types of separation simultaneously, a scenario common in real-world healthcare systems.

Methods

The proposed "confederated machine learning" method employs a three-step approach:

  1. Conditional Generative Adversarial Networks (cGANs) for Data Imputation: A central analyzer, which has a small subset of data with all types and matched IDs, trains cGANs. These networks learn to infer one data type from another (e.g., inferring medication history from diagnoses). This step is crucial because it allows for the imputation of missing data types within silos, even if a patient's complete record is not available. The cGANs use both adversarial loss (to generate realistic data) and matching loss (to ensure generated data is close to observed data).
  2. Data Inference in Silos: The trained cGAN models are distributed to the individual data silos. Each silo then uses these models to infer any missing data types based on the data it possesses. For instance, a pharmacy silo might infer diagnoses based on its medication data.
  3. Federated Training: After each silo has a more complete dataset (including imputed data), a standard federated learning approach (specifically, federated averaging) is applied. Task-specific models (e.g., for disease prediction) are trained collaboratively across all silos without moving raw data.

This method is novel in its ability to handle all three dimensions of data separation without requiring patient ID matching or complex computational infrastructure.

Datasets

The study utilizes claims data from a major U.S. health plan. This dataset includes:

  • Insurance plan type and coverage periods.
  • Age, sex, medications, and diagnoses.
  • The data covers a 48-month period, divided into an observational period, a gap period, and a test period.

For the experiments, the data was simulated to represent 99 distinct silos across 34 U.S. states, encompassing clinical silos (diagnoses), pharmacy silos (medications), and lab silos. A subset of data from California was used as the "central analyzer" with fully connected data.

Key Insights and Ideas

  • Imputation as Data Augmentation: The paper frames the imputation of missing data types using cGANs as a form of data augmentation. This is a common technique in deep learning to improve model robustness, and it appears to work effectively even with the inherent noise in healthcare data.
  • Decoupling Model Training from Data Movement: The core innovation is enabling model training without requiring data aggregation or patient ID matching, which are major obstacles in healthcare.
  • Robustness to Missing Data Types: The confederated learning approach is resilient to situations where an entire data type might be missing for some patients or even entire silos.
  • Value of Central Analyzer Size: The performance of the confederated learning model is positively correlated with the size of the central analyzer's dataset. A larger central dataset leads to better imputation models, which in turn improves the overall federated model performance.

Key Results

  • The confederated learning approach achieved competitive performance compared to a fully centralized model. For example, in predicting diabetes onset, it achieved an AUCROC of 0.787, which is close to the centralized performance of 0.824.
  • The method significantly outperformed models trained solely on data from the central analyzer or on only a single data type across silos.
  • The study demonstrated that even silos with a large number of patients can benefit from participating in confederated learning to improve their model performance.
  • The performance of the confederated model was sensitive to the size of the central analyzer's dataset, with performance plateauing around 5000 individuals in the central analyzer.

Limitations and Future Work

  • Inter-Data Type Interaction: While the cGAN imputation works well, in some applications, direct interaction between different data types might be more crucial, and this method might not fully capture it.
  • Limited Data Sources: The study primarily used claims data. More diverse datasets would be needed for more generalizable conclusions.
  • Rare Diseases: The current approach might not be sufficient for identifying very rare diseases (prevalence < 1%), where even centralized learning struggles. Future work could focus on improved representations for such cases.
  • Diagnosis Code Limitations: Using diagnosis codes for labels has limitations, as they are primarily for billing and not all silos have access to them.
  • Future Directions: The authors plan to explore other types of data separation (e.g., temporality, insurance plan) and develop meta-learning based confederated learning methods that might not require a central analyzer at all. They also aim to test the approach on more realistic silos separated by all three dimensions.

Summary updated

Abstract

Highlights•Patients' data are horizontally separated by individual and vertically separated by type.•Confederated machine learning to model data both horizontally and vertically separated.•Real-world medical data from many locations across the country.
BibTeX
@article{liu2022confederated,
  author = {Dianbo Liu and Kathe P. Fox and Griffin M. Weber and Timothy A. Miller},
  title = {{Confederated learning in healthcare: Training machine learning models using disconnected data separated by individual, data type and identity for Large-Scale health system Intelligence}},
  journal = {J. Biomed. Informatics},
  year = {2022},
  url = {https://openreview.net/forum?id=Dd0vrKcXZm},
}
Figure