Tuomo Hartonen, who once simulated the decay of Higgs bosons at CERN, will lead health AI research as a new FIMM-EMBL Group Leader starting September 1. Hartonen’s appointment follows a competitive international search and comes after he received the Academy Research Fellowship from the Research Council of Finland in June 2025, validating his interdisciplinary approach.

He uniquely combines master’s degrees in theoretical physics and translational medicine to develop multimodal health foundation models integrating diverse data sources. “I think it is important for aspiring scientists to understand that there isn’t only one path,” Hartonen emphasizes, “and that you can still create a career in science even if you don’t find your field early on.”

Multimodal Foundation Models for Population Health Trajectories

Foundation models arrange diverse health data into comprehensive trajectories, moving beyond single-purpose prediction models, according to Tuomo Hartonen. His group’s approach integrates diagnoses, medications, lab results, and medical imaging to create a complete view of patient health, rather than focusing on isolated data points. “The key idea behind foundation models is to arrange all this data into health trajectories and model it all together, rather than developing single, task-specific prediction models using only a couple of data modalities,” Hartonen explains.

This unified modeling strategy aims to improve the accuracy and scope of health predictions and interventions. A significant challenge in developing these models lies in balancing data privacy with model performance; powerful AI can inadvertently learn and perpetuate existing biases present in training data. Hartonen’s team is actively investigating this trade-off, seeking methods to enhance privacy without sacrificing accuracy or fairness.

This work extends to a Nordic-Baltic collaboration focused on evaluating and mitigating biases in health foundation models across multiple countries. Data security is paramount, with all individual-level data processed within rigorously audited cloud computing environments. While data is pseudonymized, Hartonen emphasizes the importance of secure infrastructure for handling sensitive health information and the substantial computational resources needed to train large AI models.

“The data we work with is pseudonymized,” Hartonen says. “Nevertheless, the individual-level data is only stored and processed within cloud-based computing environments that have been audited for this purpose.” He draws on his earlier experience, stating that combining this infrastructure with rich datasets and multidisciplinary expertise, FIMM aims to advance population-scale AI approaches for improved human health and foster international collaborations.

I was so impressed by the scale of that effort and how scientists there really got to push the frontiers of human knowledge.
Tuomo Hartonen, FIMM-EMBL Group Leader

Fairness and Privacy Challenges in Health AI Development

Health foundation models, while promising improved predictions, present significant hurdles in ensuring equitable and secure outcomes for diverse populations. Hartonen’s team is actively quantifying the relationship between model accuracy, fairness, and data privacy, recognizing that enhancing one often diminishes another. “Powerful AI models are very good at accidentally learning biases from the training data if such biases are present,” he explains, highlighting the risk of perpetuating existing health disparities through algorithmic outputs.

This concern extends beyond simple accuracy; the group investigates whether predictions hold consistent value across different demographic groups. Healthcare systems routinely record vast amounts of heterogeneous data from their users, creating a complex landscape for responsible AI development, FIMM-EMBL says. Protecting sensitive health information during model training is paramount, yet current privacy-enhancing techniques often come at the cost of predictive performance. Hartonen’s group seeks to map these trade-offs, aiming to identify methods that maximize both utility and confidentiality.

Hartonen reflects, drawing parallels between the rigorous pursuit of fundamental physics and the challenges of building trustworthy AI for healthcare. The team’s work isn’t solely focused on technical solutions; they also grapple with defining fairness itself. “In the context of AI models, this means are the predictions of an AI model equally good for different groups of people?” he asks, emphasizing the need for clear ethical guidelines and transparent evaluation metrics.