Introducing ourFederated Learning Platform
The vision behind the platform
An open hub for clinicians and researchers to work together defining, developing and validating AI models to improve the way we currently diagnose and treat hematological diseases
Your data stays private
Our privacy-by-design approach relies on a Federated Learning framework to ensure that no patient data ever leaves clinical sites, offering security in all data exchanges, model training and storage
An ethics-first approach
A new paradigm built on robust, ethical agreements to bring community members together through explainable AI, enabling collaborative, cross-border data sharing that is standard-compliant
Federated Learning as a key enabler
Discover the core functionalities of GenoMed4All’s Federated Learning platform regarding the Machine Learning lifecycle
Training datasets
Explore the datasets available for training FL models
Setting up new nodes
Extend the federation by including new nodes
Model validation
Validate models by running several checks when they are uploaded
Running FL tasks
Run tasks by combining models and data with common schemas
Immediate insights
Automatically gather the resulting parameters
Keep it private
Secure the federation through access control mechanisms via IAM
Model catalogue
Explore the lists of available models in the model registry
Monitoring tasks
Check task logs, queues and scheduling
Bringing together the clinical and research dimensions
We believe that accountability, transparency and usefulness of AI tools is key to build trust among healthcare professionals. Thus, we envision a clinical validation flow for these tools that successfully undergoes the meets the required standards for performance excellence in a clinical setting.
For clinicians
A local decision-support system to input new prospective and retrospective patient data, extracting insights from an ever-learning mode
For researchers
An AI workplace to explore datasets, develop, test and train new models on real-world data using our Federated Learning environment
Clinical mode
A pre-trained predictive model for clinicians and patients
How it works
1
Clinicians extract data during patient visits, possibly also through laboratory tests
2
Clinicians then input individual patient’s data into the platform and select a specific predictive algorithm based on the patient’s needs
3
The model outputs a prediction on the patient’s health status, which is interpreted by clinicians (alongside all other evidence they may have collected on the patient)
The benefits
FL provides clinicians access to predictive algorithms that can support daily practice
The FL architecture is designed in such a way that it is very natural to retrain predictive models with each new data point
FL supports the customization of predictive models for each individual patient while also incorporating global information from all other samples in the dataset
Research mode
A sandbox of retrospective (secondary) data for IT/data specialists and/or clinical data scientists and researchers
How it works
1
Clinical data scientists perform an initial exploration with local data to define the predictive model’s architecture and hyperparameters
2
Data providers upload their data to be transformed to our Common Data Model (CDM)
3
The predictive model is trained on the whole dataset, including local contributions from several different organizations via the FL paradigm
4
The global predictive model can be broadcasted back to the sources and must be tested for generalization capability and absence of bias
The benefits
FL leverages data from multiple providers with minimal degradation of model performance, thanks to its non-standard training procedure
The final models can be expected to have better generalization properties and less bias
FL enables clinical data providers to contribute to the training of predictive models while respecting patients’ privacy and GDPR regulations