CODA: an open-source platform for federated analysis and machine learning on distributed healthcare data

Area of Science:

• Health Informatics
• Computational Biology
• Medical Data Analysis

Background:

• Distributed computations offer efficient multi-institutional data analysis, avoiding data pooling complexities.
• Existing federated learning (FL) approaches lack essential features like medical standards, no-code visualizations, and robust disclosure controls.

Purpose of the Study:

• To develop and deploy the Collaborative Data Analysis (CODA) platform, enhancing distributed data analysis capabilities.
• To address key stakeholder needs by incorporating medical standards, advanced visualizations, and granular disclosure controls.
• To demonstrate the platform's utility for multi-modal federated learning using public healthcare datasets.

Main Methods:

• The CODA platform was designed and developed, incorporating features identified through stakeholder surveys.
• Multi-modal federated averaging was demonstrated using the MIMIC-IV and MIMIC-CXR datasets.
• Technical feasibility and scalability were assessed through deployment at nine Canadian hospitals, involving over one million patients.

Main Results:

• The CODA platform was successfully developed and deployed in a Canadian public healthcare setting between January 2020 and January 2023.
• Eight out of nine participating sites successfully deployed the platform, enrolling over one million patients.
• Data mapping from legacy systems to FHIR presented the primary implementation challenge.

Conclusions:

• The CODA platform provides a robust solution for distributed, multi-institutional data analysis in healthcare.
• Successful deployment in a heterogeneous IT environment demonstrates the platform's technical feasibility and scalability.
• Future work will focus on prospective model validation and tools for data format migration (FHIR/DICOM).