The rapid evolution of technological devices in biomedicine allows measuring increasingly complex and high-dimensional data (high-frequency time series, 3D images generated by medical scanners etc). Even with such availability of digital device monitors, relatively few are accepted for primary outcome measurement in late phase trials. This session explores the issues, and how they can be addressed. Among many possible approaches, functional data analysis methods provide the basic tools to extract knowledge from such intrinsically smooth biomedical data and constitute a promising frontier of research in high-dimensional statistics.

In the era of precision medicine, the paradigm of the “average population trial” is challenged by the need to adopt individualized therapies where treatment choices are dynamically linked to the course of the disease and tailored to the single patient. Biomarker discovery is critical for precision medicine because it reveals molecular features that predict response or risk with targeted treatments. In this setting, the development of modern trial designs and of computationally intensive statistical methods, e.g. for integrating multi-omics data or drug response profiles, is becoming central for predicting a patient’s prognosis or treatment response, and for accelerated drug discovery and companion diagnostics, for instance in oncology.

This session advances methods that make modern AI/ML decision-ready for biomedical research where data are censored, heterogeneous, and small in key subgroups. We showcase approaches that (i) perform causal discovery across multi-cohort data, leveraging overlap while guarding against cohort-specific biases and shifts; (ii) transfer information from large external cohorts to under-represented populations via principled weighting that aligns covariates and outcomes while down-weighting mismatched sources; and (iii) estimate individualized causal pathways with survival outcomes using deep generative models. Across these contributions, common pillars are explicit estimands, assumption transparency, and valid uncertainty with theory-backed guarantees, paired with scalable computation for longitudinal and high-dimensional settings. The importance is practical and immediate: clinicians and policymakers increasingly rely on AI trained on imperfect data. By addressing censoring, sample-size limitations, domain shift, and hidden bias, this session delivers tools that convert complex real-world data into credible, transportable causal evidence for precision medicine and public-health decisions.

AI fairness is an important topic for AI policy, and for research and healthcare. The recently published TRIPOD+AI reporting guideline and the TRIPOD AI risk of bias tool mention fairness, but do not specify how to evaluate fairness. This session starts with a discussion of the results of a systematic review of clinical predictive AI fairness metrics (https://arxiv.org/abs/2506.17035), that unveiled important shortcomings and a fragmented landscape. Next, we investigate assumptions underlying decision curve analysis that have implications for fairness evaluation. Lastly, we present a software solution to empower independent fairness evaluation while preserving intellectual property and viability of prediction models in the context of the evolving regulatory landscape.

The increasing amount of data being generated in healthcare demands new technologies for data organisation, integration, management and exploration. Semantic Web Technologies and Knowledge Graphs have proven to be suitable approaches to support health data management, and the definition of FAIR metadata improves reusability of health data records. In this session we will discuss latest developments and applications in the field.