Research Approach
Research Approach
The work progresses from responsible data collection and harmonisation to AI development, prototype integration and laboratory-scale validation.
Combining data across the patient journey
Lung cancer care produces many different types of information at different points in time. These may include:
- CT scans
- pathology images
- electronic health records
- blood-based biomarkers
- laboratory tests
- treatment information
- clinical outcomes.
LUMINA will organise these data as longitudinal patient records rather than treating each observation as an isolated snapshot. This will allow the project to examine how imaging features, biomarkers, treatments and clinical outcomes relate to one another as a patient’s condition changes.
Because health data are recorded differently across hospitals and countries, the project will establish shared structures, terminology and processing methods. The resulting harmonised data will form the foundation of the LUMINA Data Suite, which will prepare multimodal information for use by the project’s AI models.
Generating and completing multimodal data
Clinical datasets are often incomplete, and some types of information may be unavailable for individual patients. LUMINA will develop generative AI models that learn relationships between different data modalities and use them to generate realistic synthetic information.
This includes the project’s Blood-to-Image approach, which will investigate whether blood-based biomarkers can be translated into image-like representations. More broadly, the models will explore how one available data type can help estimate another that is missing.
Synthetic data will also be used to supplement the project’s clinical datasets and support controlled testing. The generated information will be evaluated for quality, clinical plausibility and consistency with real data.
Combining data-driven AI with biological knowledge
At the heart of the BRONCHI Digital Twin is a new Physics-Informed Attention Network. This methodology combines patterns learned from patient data with knowledge of lung physiology, tumour growth and treatment response.
Rather than relying entirely on statistical relationships, the models will incorporate biological and physical constraints. These constraints are intended to help the AI generate physiologically plausible predictions and make parts of its reasoning easier to interpret.
The methodology will initially be developed for non-small cell lung cancer and immunotherapy. It will then be extended through transfer learning to cover additional lung cancer subtypes and treatment strategies. The models will investigate future disease states, treatment response and patient-specific risks such as metastasis, recurrence and survival.
Supporting transparent and explainable predictions
Clinical relevance depends not only on predictive performance, but also on understanding how a model arrives at its output. LUMINA will therefore combine the inherent structure of its physics-informed models with established explainable AI methods.
These methods will help identify which imaging features, biomarkers, clinical events or model components have contributed most strongly to a prediction. The purpose is to make outputs more traceable and support expert assessment of whether they are clinically plausible.
Uncertainty will also be considered during validation so that the reliability and limitations of model outputs can be assessed alongside their predictive performance.
Developing with clinicians across four countries
Clinical and research partners in Denmark, Italy, Poland and South Korea will contribute data, medical expertise, AI development and validation capabilities. The international composition of the consortium will allow the methodology to be developed and assessed using data from different healthcare systems and patient populations.
Clinicians will contribute throughout the process, including the development of data protocols, annotation of clinical images, evaluation of synthetic data and assessment of the prototype’s predictions and explanations. Their structured feedback will be used to refine the methodology during laboratory-scale validation.