Dear Leiden University-only students, below you can read about the open thesis positions at TDS Lab for 2026-2027. More topics on these research fields are available as well. Contact me if you are interested in one of these Master's Thesis Research Projects. Cheers, Marco
Open thesis projects @TDS Lab
Security and Privacy Evaluation of a Privacy-Preserving Machine Learning Platform
Phaeton is a privacy-preserving platform that enables researchers to develop machine learning models using publicly available dummy data and subsequently submit their models for evaluation on sensitive datasets hosted within secure environments, such as those of public health organizations. Researchers receive aggregated performance reports without gaining direct access to the underlying data.
This project focuses on evaluating the privacy and security risks of this workflow. The student will investigate whether a potentially malicious participant could extract sensitive information through submitted model code, repeated evaluations, or the resulting performance reports. Through practical experiments, the student will identify possible vulnerabilities and assess measures to mitigate information leakage.
Expected outcome: An experimental security assessment of Phaeton, including identified risks and recommendations for improving the platform's privacy safeguards.
Supervisors: Marieke Vinkenoog (LUMC), Marco SpruitFine-tuning and Evaluating Decision Models for Clinical Prediction from Unstructured Text
This project aims to investigate the feasibility and performance of open source System-1 decision models (with Jev as currently best known commercial example), for predicting standardized clinical outcomes directly from unstructured medical text.
Using existing datasets from APOP and TENT studies in geriatrics, a pretrained decision model should be fine-tuned to predict clinical characteristics such as functional status, cognitive functioning, and other routinely assessed measures from referral letters, clinical notes, and discharge summaries.
These models are recent and attractive alternatives to conventional NLP architectures and large language models, as they are relatively compact, computationally efficient, and designed for fast, task-specific predictions rather than open-ended text generation. The thesis would evaluate their predictive performance, calibration, computational efficiency, and clinical applicability, ideally benchmarking them against conventional machine learning and transformer-based approaches.
Supervisors: Marco Spruit, Bram van Dijk, Simon Mooijaart (LUMC)Designing a VHT system and its experimental integration
The European Virtual Human Twins flagship Initiative under the Apply AI strategy aims to realise a Virtual Human Twin as an integrated multi-level, -time, and -discipline digital representation of a body, organ, or cell enabling the comprehensive characterisation of the physiological and the pathological state in its heterogeneity, allowing patient-specific predictions for the prevention, prediction, screening, diagnosis and treatment of a disease, as well as the evaluation, optimisation, selection, personalization, and de-risking of intervention options.
Your grand objective is to design and prototype a PoC for establishing an integrated and interoperable VHT system that couples patient-specific multimodal representations with mechanistic multiscale models through biologically interpretable and identifiable parameters and iteratively constrains and refines patient-specific VHTs using matched patient-derived experimental systems.
For inspiration and a good introductory overview on mechanistic learning approaches, see A review of mechanistic learning in mathematical oncology. This work also draws upon van Thiel's recently completed thesis Personalised and realistic training scenarios with artificial patients using AI models in Trauma Care.
Supervisor: Marco Spruit, Marcel Haas[DL] Classifying physical neighbourhouds for health impact measurement from satellite and streetview data
The correlation between health and the physical urban or rural environment remains a challenge. How do we link statistical data to the physical, built environment? How can we categorise neighbourhoods? This thesis project is part of the national ECOTIP project where we focus on identifying crucial moments when shortages of green, healthy food and clean air in neighbourhoods affect health. It is particularly the less affluent population groups that often develop lifestyle-related diseases, such as obesity, diabetes, cardiovascular diseases, mental and lung problems. At the same time, these population groups often have an unfavourable physical living environment: a lack of nature, an unhealthy food supply, air pollution and noise pollution in their neighbourhood. All these factors can affect people's healthy lifestyle. ECOTIP aims to reduce these health differences in the Netherlands.
Your task is to employ the wealth of publicly available data on the built environment, such as Google Maps, Google Earth, other satellite imagery and aerial photographs, Street View, and social media, to uncover any correlations between particular types of health indicators and the physical form of neighbourhoods where these indicators apply. For example, extract and correlate variables such as street pattern (the type of network), street width, building height, continuous development (street walls) or detached buildings (towers, apartment blocks, villas), squares, green spaces, and mobility patterns, to categorise the physical appearance of neighbourhoods based on their likely health impact.
Supervisors: Marcel Haas, Hielke Muizelaar (LUMC), Marco Spruit[LLM] Synthetic clinical notes generation in EHRs
LLMs are great in generating well-formed sentences. However, this is not desirable when we want to synthetise clinical texts that actual doctors would write: They use acronyms a lot, merely write down key phrases, some spelling errors, etc. What we want is (1) a LLM-based generator for clinical texts that produces human-like texts (2) synchronised with the tabular patient summary data. (3) It should become a module in SynthyVerse: https://synthyverse.readthedocs.io/. Inspiration can be found in https://github.com/qforge-dev/torque among others.
Daily supervisors: Marco Spruit and others
NB: A selection of former thesis topics can be found here.