Authors: Margarita Grammatikopoulou, Vasilis Alepopoulos, Giorgos Giannios, Spiros Nikolopoulos, Centre for Research and Technology Hellas.
The first deliverable of WP4, “D4.1: Initial development for decision-support and public engagement”, recently submitted, highlights the key progress and concrete results achieved during the early development of the predictive modelling and federated learning infrastructure within the PREPARE project.
The report describes the implementation of a federated learning pipeline incorporating OHDSI tools along with developed tools and solutions in the context of the project, with a federated learning simulation approach enabling the partitioning of databases to mimic distributed environments. This work supports the secure and privacy-preserving development of models across multiple institutions.
To support these activities, both synthetic datasets and real-world datasets (in addition to those from clinical partners) have been utilized, enabling robust testing and early validation. Notably, initial predictive models have been developed considering the Eunomia synthetic dataset, the real-world MIMIC III dataset and the AUTH clinical case. Both the Patient-Level Prediction (PLP) package and Deep PLP were leveraged for model development. As part of this effort, modifications were made to the existing codebase, while the Deep PLP framework was extended to incorporate additional algorithms, to meet the specific requirements of the project. These models are currently being used to guide further development and case-specific refinement.
An interactive model exploration tool, inspired by OHDSI’s PLP viewer, has been built to facilitate internal review and communication of results, both within the project and as part of broader public engagement strategies.
Furthermore, mock-ups of graphical user interfaces (GUIs) for patient stratification and scenario building tools have been produced, laying the groundwork for the PREPARE platform’s components for the clinicians.
From a deployment perspective, the use case of the clinical partner AUTH has been executed end-to-end, while initial steps have begun for the clinical use case of ISICO, with RAD and EMC planned to follow. Extensive testing from available synthetic and real-word datasets has yielded valuable insights and lessons learned, which will directly inform the ongoing development of models in subsequent cases.
In parallel, bias and fairness considerations have been actively integrated into all activities within WP4. A dedicated bias and fairness assessment tool has been developed and embedded into the federated learning simulation pipeline, ensuring that equity and transparency remain central to model evaluation.
These achievements establish a strong foundation for the next phases of model development and case expansion. Additionally, initial steps have been made towards defining the public engagement strategy, particularly focusing on engagement with the technology research community to support model dissemination and validation.
The activities within the PREPARE project towards model development highlight the importance of a multidisciplinary approach, emphasizing the need for continuous collaboration between technical and clinical partners. The success of predictive and stratification models is based on both the technical robustness of the model architecture and the semantic clarity and completeness of the mapped data. Continuous refinement of the process, transparent documentation and collaboration between consortium partners is essential in ensuring clinically relevant and interoperable modelling outcomes.
Outcomes of WP4 activities within the PREPARE project:
- A federated learning pipeline has been established for model development
- A federated learning simulation approach is in place to simulate the participation of several clinical sites in model training, to test the prediction model aggregation and evaluate the resulting “federated global model”
- Synthetic and real-world datasets were leveraged to ensure progress without delays, while clinical partners’ real-world datasets will be utilized next
- Extensive testing has led to valuable lessons learned, which will guide model development across all clinical partner use cases
- Initial predictive models are available for the AUTH use case
- An interactive model exploration tool – similar to OHDSI’s PLP viewer – was developed to visualize, review and communicate model results within the project and for public engagement
- An initial public engagement strategy has been defined, along with corresponding outreach activities
- Mock-ups of GUIs for patient stratification tools and scenario building tools have been established for four clinical partners’ use cases
- Bias and fairness activities are actively supporting and monitoring all activities in WP4
- A bias and fairness assessment tool has been developed and integrated into the federated learning simulation pipeline
.
Posted by: Nayara Güércio (Trilateral Research)