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 (horizontally) but also by data type (vertically) and patient identity (making direct matching impossible). This method allows for robust model training without centralizing sensitive patient data or requiring difficult patient ID matching, paving the way for more powerful AI in healthcare.
What Problem Does This Paper Address?
Healthcare data is notoriously siloed. Patient information is scattered across different providers, hospitals, and even within different departments of the same institution. This fragmentation makes it incredibly challenging to build comprehensive AI models that can predict diseases or identify health risks. Traditional federated learning addresses data separated horizontally (different patients in different locations), but it struggles when data for the same patient is split vertically (e.g., diagnoses in one system, medication records in another) and when patient IDs are inconsistent or unavailable across these systems. This "separation by identity" is a significant hurdle. The authors aim to overcome these multi-dimensional data separations to enable effective machine learning for large-scale health system intelligence.
Methods
The core of the paper is the proposed "confederated machine learning" method, which operates in three steps:
- Conditional Generative Adversarial Networks (cGANs) for Data Imputation: The central analyzer (a hypothetical entity with access to a small, fully connected dataset) trains cGANs. These networks learn to infer one type of data from another (e.g., inferring medication data from diagnosis data). This is crucial because not all patients will have all data types available in every silo. The cGANs use both adversarial loss (to generate realistic data) and matching loss (to ensure generated data is close to observed data).
- Data Inference in Silos: The trained cGANs and a disease classifier are sent to each of the distributed data silos. Each silo then uses these models to infer missing data types for its patients. This step effectively "fills in the gaps" for each silo's local data.
- Federated Training: After data inference, each silo has a more complete dataset. A standard federated learning approach (specifically, federated averaging) is then applied to train task-specific models (e.g., disease prediction models) across all silos simultaneously, without sharing raw data.
This approach is novel in its combination of data imputation via cGANs with federated learning to handle multi-dimensional data separation, particularly the "separation by identity" challenge.
Datasets
The study uses claims data from a major U.S. health plan, covering July 2016 to June 2019. The dataset includes insurance plan type, coverage periods, age, sex, medications (represented by National Drug Codes - NDC), and diagnoses (represented by ICD-10 codes). Lab test data was not available for this study.
The data was simulated to represent a federated network of 99 silos across 34 states. A single state (California) was designated as the "central analyzer" with access to fully connected data for 5433 individuals. The remaining data from 33 states was divided into 99 silos, simulating horizontal separation (by state), vertical separation (by data type: diagnosis, pharmacy, lab), and separation by identity (no patient ID matching).
Key Insights and Ideas
- Imputation as Data Augmentation: The paper frames the cGAN-based data imputation not just as filling missing values, but as a form of data augmentation, similar to techniques used in computer vision to make models more robust. This is a key insight into why the method can tolerate potential noise from imputation.
- Decoupling Data Types and Identity: The core innovation is enabling federated learning even when patient IDs are not reliably matched across silos. This bypasses a major practical barrier in healthcare data sharing.
- Central Analyzer's Role: The central analyzer doesn't need a massive dataset, but it does need a representative sample with complete data to effectively train the imputation models. The size of this central dataset influences the imputation quality and, consequently, the overall model performance.
- Robustness to Missing Data Types: The method is designed to handle situations where an entire data type might be missing for some patients or even entire silos, making it more practical for real-world healthcare scenarios.
Key Results
The confederated learning approach achieved competitive performance in predicting the risk of diabetes, psychological disorders, and ischemic heart disease. For instance, in predicting diabetes onset, the confederated learning model achieved an AUCROC of 0.787, which is close to the performance of a fully centralized model (0.824) and significantly better than models trained only on data from the central analyzer (0.774) or federated learning using only diagnosis data (0.775). Similar trends were observed for other diseases.
The study also found that the performance of the confederated model generally improved with a larger number of people in the central analyzer's dataset, suggesting an incentive for larger institutions to participate.
Limitations and Future Work
- Central Analyzer Dependency: The method relies on a central analyzer with some fully connected data. While this is a simplification, future work aims to explore meta-learning approaches that might eliminate the need for a central analyzer altogether.
- Limited Data Types and Interactions: The study primarily used diagnoses, medications, and lab tests. More complex inter-data type interactions might be crucial for other applications, and the current method might not fully capture them.
- Rare Diseases: The current approach showed limitations for very rare diseases, where even centralized learning struggled. Future work will focus on improving representations for classifying rare diseases in a federated setting.
- Labeling Bias: Using diagnosis codes for disease labels has limitations, as these codes are often used for billing and may not perfectly reflect clinical diagnoses. Additionally, silos like pharmacies and labs may not have access to diagnosis codes, limiting their ability to train models solely on their data.
- Generalizability: The study used a simulated dataset. Future work will involve testing the method on more realistic, diverse silo configurations and exploring other types of data separation (e.g., temporality, insurance plan).
Summary updated