The proposed national AI strategy report highlights the promotion of health care for all as one of the top priorities for AI in Greece, bolstering the national health system with the capacities to provide better quality, data driven, and targeted health care, including the prediction and management of chronic and rare diseases. The country is currently undergoing a major digital transformation effort in its public healthcare system (incl. investing in digitizing patient data, such as prescriptions and examinations) to build a national Electronic Health Record (EHR) system, which creates a leapfrogging opportunity for the collection of AI-ready data at a population-wide level. The goal of existing projects is to digitize national healthcare system records and connect them with private practices nationwide, thereby establishing a rich data ecosystem for AI applications. This will position the country at the forefront of healthcare innovation, enhancing the quality and efficiency of patient care, and setting a global standard in healthcare technology. At the same time, such efforts will create a vast, trusted, anonymized and secure data pool, which is crucial for training and developing AI models. This rich data repository can become a valuable asset for universities, research institutions, start-ups, technology companies, and pharmaceutical companies, fostering collaborations between academia and industry. This competitive edge could attract foreign investment as global entities seek to leverage this data for ground-breaking research and development in the field of healthcare, including chronic diseases. This influx of investment can spur economic growth, fund further medical advancements, and establish Greece as a hub for medical research and technology, resulting in better healthcare for the Greek people.
Pharos AI Factory intends to establish and operate a dedicated Health AI Hub, in order to support and boost health-related aspirations. In particular, the objectives of Pharos AI Factory in the field of healthcare will be:
• The leveraging of AI / HPC / HPDA tools and services for exploiting the existing national and international health data repositories (descriptive statistics, predictive modeling of big datasets, training deep learning models on medical imaging datasets, etc.)
• The transparent and human-centered integration of Generative AI and Large Models (such as LLMs, VLMs) in Health Services
• The creation of trust among citizens with the use of responsible AI, efficient and friendly use of services, explainability, respect of privacy, taking into consideration the AI act and Data Act.
• The creation of an Innovation Ecosystem based on AI with synergies from Public Sector, Academy, SMEs, startups, Large companies, Centers of Excellence, Competence Centers, as well as Digital Innovation Hubs (eDIHs).
Main outcomes of the hub will be:
• Develop data management pipelines (anonymization, pre-processing, harmonization, curation, anonymization, annotation etc.) to effectively support data preparation for AI-ready formats, aligned with health data standards and ethical guidelines
• Provide generative AI models for synthetic data generation in domains wherethat data availability is limited (e.g. costly experiments, low-throughput diagnostics, rare diseases, etc), ensuring that models can be effectively trained without compromising on data integrity or patient privacy
• Implement a centralized repository consisting of AI-ready datasets, AI models, and machine learning code, using version control. The repository will support common data formats such as CSV and JSON, and model formats such as PyTorch and TensorFlow, ensuring broad compatibility. A web interface will allow users to browse and search with metadata tagging easily. Access controls will manage permissions for users, ensuring that only authorized ones can access and/or modify data. Audit logs will track changes, providing detailed records of who accessed or altered data, enhancing security and compliance in a collaborative research environment.
• Develop, train, test, and optimize AI models on the HPC resources with improved performance and scalability. Using distributed computing frameworks such as TensorFlow, PyTorch, and Horovod, we will accelerate model training and enable the processing of large datasets, providing the model and code as deliverables (?).
• Streamline model development, by integrating AutoML (Automated Machine Learning) capabilities, allowing users to automate the selection of algorithms, hyperparameter tuning, and feature engineering. This will empower researchers and developers to efficiently explore numerous model configurations, leading to optimized performance with minimal manual intervention.
Indicative models
• AI models on multimodal biomedical data (e.g. imaging, sequencing, medical records, open datasets, etc) to enable a holistic approach to patient diagnostics and treatment
• domain-specific AI models (e.g. functional annotation of genomic variation in breast cancer, AI-mediated clinical assistance based on magnetic resonance imaging of patients with Alzheimer’s disease etc.) to support targeted healthcare solutions
• AI prediction models for disease progression, (focusing on large scale visual language models for medical imaging, as well as on survival machine learning models for Electronic Health Records), including segmentation, classification, fair and explainable decision making
• Design sophisticated interpretability tools incorporating intelligent human interfaces to offer patients, clinicians, informal caregivers, and policymakers, interpretable and actionable insights into AI-driven decisions. These tools will be customised to provide tailored explanations according to each end user’s needs.
In this context, enhanced visualisations will be combined with natural language explanations, enabling the end users to comprehend the models’ outcomes, thus fostering trust and supporting informed decision-making.
– Integrate uncertainty quantification mechanisms into intelligent clinical decision interfaces, based on Bayesian methods, Monte Carlo dropout, and ensemble modeling. By visualizing uncertainty metrics (e.g., prediction intervals, variance heatmaps) alongside AI predictions, these interfaces will help end users assess the confidence of AI outputs in real time, thus improving decision safety and fairness in healthcare applications.
– Implement drift monitoring and out-of-distribution detection tools for the ongoing assessment of model performance and the identification of data drifts in real-time. In this context, end users will be provided with real-time alerts, in case model performance degradation and/or deviations from expected data patterns are detected, along with actionable recommendations (e.g., human review, initiation of appropriate adaptation processes), with the goal to facilitate the reliability and robustness of the developed AI models.
– Develop continual learning and model transportability pipelines towards ensuring the development of adaptive AI models, that will evolve with new data while maintaining their performance across diverse environments and over time. Domain adaptation approaches and transfer learning methods will be deployed to ensure timely model updates and refinements in response to changing data distributions, new clinical contexts, and varying patient demographics.
Based on the above, indicative applications include:
• AI tools to assist the transition from basic research to clinical practice (e.g. from AI-driven modeling of single-cell and spatially resolved transcriptomics to understanding resistance to therapeutic compounds)
• learning methods to leverage knowledge from well-explored domains and gain insights in areas with low data availability, enhancing the accuracy of predictions for underrepresented health conditions
• explainability and fairness metrics in AI models to foster trust among healthcare providers and patients, ensuring responsible AI practices.
To fulfill its objectives, Pharos-Health AI Hub will tap into and closely collaborate with key European or national initiatives, such as the Digital Europe Programme (DEP)’s European Digital Innovation Hubs (EDIHs), the Greek Health Data Access Body Infrastructure (GR-HDAB) and the health-oriented European Centers of Excellence (Bioexcel & CompBioMed).
