Open Topics [MSc thesis, Internship, BSc thesis]
I am currently offering several topics for MSc theses and internships - all in the areas of machine learning, data science, sports, health, and digital technologies.
Please note that all topics can be tailored to mutually agreed-upon interests.
In-context Learning for Health Prediction Using Tabular Foundation Models
Suitable for MSc Thesis / InternshipAreas: Machine learning · Tabular foundation models · In-context learning · Prompt engineering · Zero-shot learning · Few-shot learning · Benchmarking
Summary
Large Language Models (LLMs) are increasingly becoming important tools in a wide range of applications. One of their most interesting capabilities is in-context learning, which enables models to perform new tasks by providing instructions and/or examples directly within the input, without requiring explicit model retraining. The success of LLMs has also contributed to a broader development of foundation models beyond natural language processing. In particular, recent research has introduced foundation models specifically designed for tabular data, with promising results for prediction and classification tasks. Some early studies suggest that these models can achieve competitive, and in some settings superior, performance compared with conventional machine learning approaches, including tree-based methods that have traditionally dominated tabular data analysis. We will investigate the potential of tabular foundation models for health prediction tasks. The project will evaluate how well these models perform across different health-related datasets and prediction problems and compare their performance with established machine learning baselines. Depending on the student's interests, the thesis may also explore zero-shot and few-shot learning, in-context learning, example selection, model prompting, and the robustness and generalisability of tabular foundation models across different health datasets.
Relevant Literature
Comparing LLMs and Tabular Foundation Models for In-Context Learning in Health
Suitable for MSc Thesis / InternshipAreas: Machine Learning · Data Science · Foundation Models · Tabular Data · Tabular foundation models · In-context learning · Prompt engineering · LLMs
Summary
Large Language Models (LLMs) are increasingly becoming important tools across a wide range of applications. Many have explored how LLMs can be adapted to perform tasks for which they were not specifically trained, using approaches such as fine-tuning and various prompting and prompt-steering techniques. These approaches can enable LLMs to adapt to new tasks without developing a task-specific model from scratch. One very interesting approach is their capability to do in-context learning. In parallel, tabular foundation models have emerged with similar capabilities, enabling in-context learning for structured tabular data. This development is particularly interesting for health research, where labelled datasets are often small to moderate in size. Training task-specific machine learning models, such as tree-based methods, support vector machines, or deep learning models, may be challenging or inefficient when the amount of labelled data is limited. we will investigate (and compare) how effectively in-context learning with emerging foundation models, including LLMs and tabular foundation models, can be used for health prediction and assessment tasks. The performance of these approaches will be evaluated and compared with established baseline machine learning models. Depending on the joint interests, the project can further explore different prompting strategies, model types, data representations, and the robustness and generalisability of these approaches across different health datasets.
Relevant Literature
Activity recognition from wearable accelerometers using pre-trained foundation models
Suitable for MSc Thesis / InternshipAreas: Activity recognition · Accelerometers · Machine Learning · Data Science · Foundation Models · foundation models · In-context learning
Summary
In epidemiological and observational research, monitoring physical activity behavior using wearable devices has become increasingly common, providing an objective alternative to traditional self-reported measures. Among wearable devices, accelerometers are one of the most widely used sensors for monitoring human movement. Accelerometer signals (i.e., motion signals) can be used to characterize different types of physical activity and quantify activity intensity, duration, and patterns under free-living conditions. To date, hundreds of algorithms and methods have been proposed, developed, and validated to transform accelerometer signals into meaningful measures of physical activity, including activity types, intensity, duration, and energy expenditure. These approaches range from traditional signal-processing and rule-based methods to modern machine learning and deep learning techniques. Although many studies have demonstrated that machine learning and deep learning models can outperform simpler traditional approaches in activity recognition and classification tasks, traditional methods remain widely used in applied health and epidemiological research. One reason may be concerns about the robustness and generalizability of machine learning models, particularly when they are developed and evaluated using small- to moderate-sized datasets. Models trained on limited and potentially heterogeneous samples may not generalize well to new populations, devices, or free-living conditions. Another potential reason is that health researchers may have limited familiarity with machine learning methods, including model development, training, validation, and interpretation. We will investigate whether pre-trained foundation models can be used for wearable accelerometer-based physical activity classification and whether they perform as well as or better than models trained on small datasets. The study will provide insights into the potential of pre-trained models to reduce the need for task-specific model development and improve the accessibility and generalizability of machine learning approaches in wearable-based physical activity research.
Relevant Literature
Daily step counts and health outcomes
Suitable for MSc Thesis / InternshipAreas: Step counts · Accelerometers · Data Science · Health · Physical activity ·
Summary
Physical activity has numerous health benefits, but promoting sufficient physical activity remains a challenge. Step count is a simple measure of daily physical activity and provides an easy-to-understand way for people to monitor their activity levels. Earlier, the goal of 10,000 steps per day eas widely promoted as a target for healthy living, despite limited scientific evidence supporting it as a specific threshold for health. More recent research suggests that health benefits can be observed at substantially lower step counts and that the association between daily steps and health is generally non-linear. Although taking more steps is generally associated with greater health benefits, the additional benefit tends to diminish at higher levels of activity. Still, more research is needed to understand how step count relates to different health outcomes and to identify whether there are meaningful step-count ranges that may be most beneficial for different populations and health conditions. Here, we will examine how daily step counts are associated with different health outcomes while accounting for overall physical activity levels. The health outcomes of interest will primarily include cardiometabolic risk markers or mental health outcomes.