Dorina Thanou is a senior researcher and lecturer at EPFL working at the intersection of artificial intelligence, engineering, and medicine. She leads strategic research initiatives in AI for Health at the EPFL AI Center and is a Research and Teaching Associate in EPFL’s School of Engineering.

Her research develops AI methods for understanding complex biological and medical systems from heterogeneous, incomplete, and longitudinal data. Building on foundations in signal processing, graph-based learning, and geometric deep learning, her work focuses on structured, multimodal, interpretable, and mechanistic learning, with the broader goal of understanding how individual health evolves across biological scales and over time. Her research spans key application domains, including precision oncology, preventive cardiology, and longitudinal health.

Dorina received her PhD in Electrical Engineering and MSc in Communication Systems from EPFL, following a Diploma in Electrical and Computer Engineering from the University of Patras. Her research trajectory has evolved from fundamental aspects of signal processing and learning on graphs toward computational approaches designed around the structure, dynamics, and multimodal nature of biological and medical systems. Prior to her current appointments, she held research positions at the Swiss Data Science Center (SDSC) and EPFL’s Center for Intelligent Systems (CIS), where she led the Intelligent Systems for Medicine and Health program, and was previously a research intern at Microsoft Research (MSR).

She leads and contributes to large interdisciplinary initiatives connecting fundamental AI methodology with clinical and biological questions. She is Executive Director and co-PI of the National AI Initiative for Precision Oncology (NAIPO) and co-PI of SwissCardIA, a large-scale prospective initiative in cardiovascular prevention integrating imaging, continuous sensing, biomarkers, multi-omics, and lifestyle data. She is an ELLIS Scholar and an IEEE Senior Member.

Human health emerges from interacting processes across multiple biological scales, from molecular and cellular mechanisms to tissue organization, organ function, physiology, and behavior. Yet our observations of these processes are inherently heterogeneous and incomplete.

Our research develops AI methods to learn from these complex observations. We seek to build computational representations that capture the structure of biological systems, integrate complementary sources of information, and model how individual health evolves over time.

Our methodological research is grounded in close collaborations with clinicians and biomedical researchers, whose domain expertise, data, and scientific questions provide the real-world context in which these methods are developed and evaluated.

Biological systems have rich relational and hierarchical structure: cells interact within tissues, tissues form organs, and observations at different scales provide complementary views of the same underlying system. Standard representations often fail to capture these dependencies explicitly.

We develop fundamental graph-based, geometric, and generative learning methods that exploit this structure and connect information across scales. Our work investigates how relational, spatial, and topological organization can be incorporated into learned representations, from cellular organization in tissues to anatomical and patient-level representations.

Biomedical systems are observed through heterogeneous modalities, each capturing different aspects of the underlying biological state. These observations may differ in resolution and information content and are often only partially available, making their integration a fundamental machine learning challenge.

We develop multimodal representation-learning methods that disentangle shared and modality-specific information, learn from incomplete combinations of observations, and incorporate prior knowledge into the learning process. We are particularly interested in using anatomical, physical, and biological structure as inductive biases to improve representation learning and reasoning from limited observations.

Health and disease are dynamic processes, yet the observations available to us are typically sparse, irregular, and asynchronous. Learning from such data requires models that go beyond individual snapshots and capture the latent processes governing how a system changes over time.

We develop structured and generative models for longitudinal and spatiotemporal data, with a focus on learning latent dynamics, identifying changes between dynamical regimes, and building representations that evolve as new observations become available. In biomedical applications, these methods enable the characterization of individual trajectories associated with disease progression, treatment response, and recovery.

Keywords: Longitudinal modeling · Spatiotemporal learning · Generative state-space models · Dynamic graphs · Predictive digital twins

Biomedical Applications & Collaborations

Our methodological research is developed and evaluated through close collaborations with clinical and biomedical partners, with a particular focus on precision oncology (NAIPO), preventive cardiology (SwissCardIA), and longitudinal health (LETITIA, PERISCOPE). These collaborations ground the development of new AI methodology in clinically relevant problems and provide access to rich, complementary observations across biological scales and over time.

Keywords: Digital pathology · Spatial biology · Medical imaging · Multi-omics · Molecular and clinical data · Biomarkers · Physiological signals · Wearable sensing · Longitudinal patient data

Please check our complete list of publications on: https://scholar.google.com/citations?user=UIHKfY0AAAAJ

My research has been supported by competitive national and institutional funding across AI methodology and its applications in precision oncology, cardiovascular medicine, medical imaging, and digital pathology. These projects support the development of new AI methods and their evaluation through interdisciplinary collaborations with clinical and biomedical partners.

A national multi-institutional initiative bringing together AI research, clinical oncology, healthcare infrastructure, and industry to advance AI-enabled precision oncology.

AI methods for longitudinal phenotyping and prognosis prediction in brain metastases treated with stereotactic radiosurgery.

Development of explainable AI methods leveraging vision-language foundation models for medical imaging.

Development of representation-learning methods for disentangling complementary factors in complex graph-structured biomedical data.

Development of AI methods for learning longitudinal tumor dynamics and early markers of treatment response from PET/CT imaging.

Translation of computational imaging and AI methods for digital pathology toward clinical use.

Development of AI methodology for cardiovascular imaging and patient-specific prediction in collaboration with clinical partners.

Development of AI approaches for digital pathology and precision oncology, with a focus on computational representations of tumor tissue.