 methodology. PHT is a comprehensive manifesto that outlines the technological, governance, and procedural elements required to implement a fully functional FL infrastructure. This approach aims to facilitate the generation of statistical models and insights without exchanging identifiable patient data among cooperating institutions.){kind=link}
A recent study, published by the British Institute of Radiology and led by a consortium of researchers, showcases the Personal Health Train (PHT) methodology. PHT is a comprehensive manifesto that outlines the technological, governance, and procedural elements required to implement a fully functional FL infrastructure. This approach aims to facilitate the generation of statistical models and insights without exchanging identifiable patient data among cooperating institutions.
The PHT framework emphasizes compliance with privacy regulations like the General Data Protection Regulation (GDPR) and the fair allocation of intellectual property rights among collaborators. To this end, the consortium has made legal agreement templates openly accessible to streamline partnerships.
Technologically, the PHT framework is built on three core components:
- Tracks: Secure telecommunications links for message transmission between a central model aggregation hub and participating institutions.
- Trains: Containerized software applications carrying data query filters and analysis code, executed locally within institutional firewalls.
- Stations: Institutional data repositories holding patient data for the creation of a single, global multi-institutional model.
Significantly, the PHT framework calls for the preparation and exploration of data to adhere to FAIR (Findable, Accessible, Interoperable, Reusable) principles before any modelling or analysis. These principles ensure that data is usable by both machine algorithms and human operators, fostering greater autonomy and interoperability among researchers and clinicians.
The research highlights the creation of a federated dashboard for data exploration and visualization, allowing for the summarization of case-mix variables from diverse datasets. This dashboard enables clinicians and researchers to select consistent cohorts from each other’s data, enhancing collaborative efforts.
To demonstrate the clinical relevance of their work, the researchers applied the PHT framework to a use case in head and neck cancer outcomes prognostication. Using a mix of public open data and private institutional data, they successfully trained a range of prognostic models based on clinical factors and radiomic features.
For more details, the full research article can be accessed at BRJ Artificial Intelligence.