The AISym4MED project has successfully transitioned from conceptual design to a functional, federated prototype, defining a distributed architecture that supports secure health data collection, aggregation, and analysis. By implementing a backend based on modular containerization, the project has established a secure environment where isolated code execution mitigates adversarial risks and ensures data integrity. Simultaneously, a human-centred frontend was deployed, aiming to make complex functionalities, such as synthetic data generation and model auditing, intuitive for diverse stakeholders, including bioengineers and clinicians. This is supported by a cross-border health database architecture using standardised schemas and clinical terminology, enabling seamless integration across diverse datasets.
Data auditing is anchored in pyMDMA (Multimodal Data Metrics for Auditing) library, providing a standardised framework for validating both real and synthetic medical images, time series, and tabular data. Regarding model auditing, new modules for explainability, fairness and bias mitigation, uncertainty estimation, and privacy have been implemented to quantify and correct model performance across protected demographic groups, assess confidence in predictions, and evaluate privacy risks, ensuring the development of trustworthy AI. GASTeN framework for stress-testing models against edge-case data was conceptualised.
Regarding synthetic medical data generation, the project has moved beyond preliminary testing to deliver high-fidelity generative models across multiple modalities and targeting the project’s use cases. Key breakthroughs include a controllable model for retinal fundus images and specific generative models for ECG, EEG, and clinical tabular data. A quantitative approach to evaluating synthetic data was proposed, unifying the dimensions of fidelity, diversity, privacy, and utility. Specific additional metrics to evaluate synthetic clinical time series were conceptualised. To bridge the gap between quantitative metrics and clinical trust, the project launched the "Doctor-in-the-Loop" evaluation workflow.
To ensure trustworthiness and data privacy, the project evolved its framework into a functional, multi-layered security architecture, based on robust legal foundation for cross-border processing.
To iterative validation of functionalities was set. Early pairing of technical partners with use-case owners, yielded the essential "building blocks" for validation, including data dictionaries, feature extraction methods, and predictive and generative models tailored to real-world clinical scenarios.
