Health and life sciences

Health and life-science data is often heterogeneous, high-dimensional, and collected under practical rather than experimental conditions. This creates fundamental questions about interpretability, uncertainty, causal inference, and the transfer of conclusions between people and environments.

MyMicroZoo: causal microbiome analysis

We collaborate with MyMicroZoo on causal analysis of gut microbiome data. Microbiome measurements describe the composition and relative abundance of many interacting organisms, while health, diet, lifestyle, and other factors may all influence the observations. Our aim is to move beyond associations by developing methods that help distinguish possible causal relationships from confounding and reveal effects that are both statistically reliable and biologically meaningful.

LUMC Neurology and the Leiden Headache Center: migraine data analysis

Our collaboration on migraine research uses longitudinal data such as electronic headache diaries and measurements collected over time. We investigate how migraine days and attacks can be characterised, which personal factors and triggers are informative, and how predictions can be made sufficiently reliable and interpretable to support research and, ultimately, personalised care.

LUMC Intensive Care: critical-care data analysis

Intensive-care data combines many measurements, interventions, and clinical events for a highly heterogeneous patient population. Together with the LUMC ICU, we study how interpretable models can identify clinically meaningful patient subgroups and risks. Recent work on ICU readmission illustrates this approach: a compact set of probabilistic rules identifies subgroups with different readmission risks and characterises them in terms clinicians can directly inspect.

Selected recent work

	Yang, L, van der Meijden, S, Arbous, S & van Leeuwen, M Interpretable Machine Learning for Identifying ICU Readmission Risk in Subgroups with Probabilistic Rules. Journal of the American Medical Informatics Association vol.33(3), pp 690-699, Oxford Journals , 2026.
	van der Arend, B, Verhagen, I, van Leeuwen, M, van der Arend, M, van Casteren, D & Terwindt, G Defining migraine days, based on longitudinal E-diary data. Cephalalgia vol.43(5), 2023.

People and society

Data about people is shaped by context: behaviour changes over time, social relationships affect observations, and measurements from diaries, sensors, or wearables provide only a partial view. These settings inspire research on finding robust behavioural patterns, modelling interactions and change, combining different types of observations, and keeping the resulting models understandable to domain experts and the people represented by the data.

Our recent work includes wearable sensing for human activity monitoring, estimating physical activity among nursing-home residents with dementia, and analysing face-to-face interaction networks among children. Such studies connect methodological questions about temporal and network data to questions about health, wellbeing, and social behaviour.

Selected recent work

	Gawehns, D & van Leeuwen, M Social Fluidity in Children's Face-to-Face Interaction Networks. In: Proceedings of the 17th International Conference on Complex Networks (CompleNet) 2026, 2026.
	Gawehns, D, Portegijs, S, van Beek, APA & van Leeuwen, M Using consumer wearables to estimate physical activity of nursing home residents with dementia. Exploration of Digital Health Technologies vol.3, Open Exploration, 2025.
	van Dijk, R, Gawehns, D & van Leeuwen, M WEARDA: recording wearable sensor data for human activity monitoring. Journal of Open Research Software vol.11(1), 2023.

High-tech industry and engineering

High-tech industrial systems generate large volumes of sensor, log, and process data, but abnormal behaviour and failures are often rare and operating conditions continually change. These applications motivate research on robust anomaly detection, explanation, predictive maintenance, learning from heterogeneous data, and the integration of data-driven models with engineering and process knowledge.

Digital Twin Perspective programme

Our group has primarily contributed to three industrial use cases in the Digital Twin Perspective programme:

• HIsarna process stability: understanding and improving the stability of an energy-efficient ironmaking process as it moves towards commercial operation;
• ASML and VDL wafer-handler robots: using health monitoring and preventive maintenance to reduce unscheduled downtime in complex mechatronic equipment; and
• Canon high-end printers: monitoring system health and enabling predictive maintenance to improve uptime and reduce service costs.

Across these cases, a useful model must do more than raise an alarm. Engineers need to understand which signals and process conditions matter, how confident the model is, and whether an apparent anomaly reflects a fault, a changing operating regime, or normal variation. These needs have directly inspired our research on trustworthy and explainable anomaly detection.

Related recent work

	Li, Z, Huang, Q, Zhu, Y, Yang, L, Mohammadi Amiri, M, van Stein, N & van Leeuwen, M Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching. In: Proceedings of the Conference on Neural Information Processing Systems (NeurIPS 2025), 2025.
	Li, Z Trustworthy anomaly detection for smart manufacturing. PhD thesis, Leiden University, 2025.