MLG proposes the following MA thesis topics for this academic year. NB: Number of topics is limited. If interested please contact the supervisor asap.

Causal Discovery in ICU Data: Leveraging Missingness as a Proxy for Latent Patient Acuity

Supervisors: Dr Gian Marco Paldino and Prof Gianluca Bontempi

Intensive Care Unit (ICU) datasets are highly complex, characterized by irregular clinical measurements and substantial missing values. Traditional statistical and machine learning approaches typically view missing data as an obstacle, attempting to resolve it through imputation or deletion. However, in critical care, missingness is rarely random. The clinician’s decision to order—or not order—a diagnostic test is a conscious choice reflecting their implicit evaluation of a patient’s health status. Consequently, these missingness patterns act as an informative proxy for latent patient acuity.

This thesis project aims to investigate and develop causal discovery methods that explicitly leverage informative missingness to reconstruct underlying causal relationships in clinical settings. The student will explore how missingness graphs (m-graphs) and constraint-based or continuous causal discovery algorithms (such as MVPC or MissDAG) can be adapted to treat clinical decision-making and missingness indicators as active nodes in the causal graph. The proposed framework will be developed in Python and validated using real-world EHR databases such as MIMIC-IV or the eICU Collaborative Research Database.

The ideal candidate should have a strong background in Machine Learning, Probability, and Statistics, be proficient in Python programming, and have an interest in causal inference and healthcare applications.

Resources & References:

Enhancing Digital Twins with Vector Databases for Real-Time Intelligent Analytics

Supervisor: Gianluca Bontempi

Digital twins are virtual representations of physical systems that enable monitoring, simulation, and optimization of real-world operations. As digital twins increasingly rely on large volumes of sensor and operational data, efficient data management becomes a critical challenge. Vector databases, designed for storing and retrieving high-dimensional data generated by machine learning models, offer new opportunities for real-time analytics, similarity search, anomaly detection, and predictive maintenance.

This thesis aims to investigate the integration of vector databases into digital twin architectures to improve data retrieval performance and support intelligent decision-making. A prototype digital twin system will be developed in Python, leveraging machine learning techniques and vector search capabilities. The proposed approach will be evaluated through a practical case study involving a dynamic system and/or industrial process.

The ideal candidate should be proficient in Python programming, enrolled in the Master’s module on Computational Intelligence, have a strong background in Machine Learning, and be motivated by interdisciplinary applied research.

Python-Based Vector Databases & References:

AI-Based Multi-Source Air Quality Prediction Using Satellite and Ground Observations

Supervisors: Ihsane Gryech

This thesis aims to develop an artificial intelligence framework capable of predicting urban air pollution by integrating heterogeneous environmental data sources. The study will combine ground air-quality monitoring station data (sourced from Airparif), satellite Earth-observation data (Sentinel-5P/TROPOMI), and meteorological variables such as temperature, humidity, and wind speed. The objective is to generate high-resolution predictions of pollutants (PM₂.₅, PM₁₀, NO₂, O₃) using machine-learning models capable of capturing complex spatio-temporal patterns.

The research will compare several algorithms, including traditional ensemble methods (Random Forest, XGBoost) and hybrid deep-learning architectures (e.g., CNN-LSTM), where convolutional layers extract spatial features from satellite imagery while LSTM networks model temporal pollution dynamics. The final outcome will be an AI-based prediction system capable of producing short-term pollution forecasts (24–72 hours) and identifying pollution hotspots, contributing to improved urban sustainability strategies.
The student should be proficient in Python and machine learning, with a strong interest in environmental data science and spatial modeling.

Industry-Ready Tools & Resources:

The use of neural networks in best basis selection for image compression

Supervisor: Maarten Jansen

This project aims at investigating the state-of-the-art in image compression (or image processing) using best basis selection using machine learning.

State-of-the-art in wavelet based image compression or image processing

Supervisor: Maarten Jansen

Find two recent and related papers (less than 5 years since publication) describing new advances in wavelet based image processing (edge detection, edge enhancement, noise reduction) or image compression. Compare the two methods and apply to test data.

Wavelet based interpretation of deep neural networks

Supervisor: Maarten Jansen

Explore at least two methods, comparing their focusses, objectives, methods and results for the interpretation of trained neural networks using wavelet transforms. For this project, the idea is not to use the wavelet decomposition as a tool in the training phase, but for the analysis afterwards.

Generative State Steering: Deep Learning for Bioprocess Optimization

Supervisors: Raymond Lochner and Matthieu Defrance

This research project aims to develop a novel deep learning framework to solve a complex “State Steering” problem, modeling a biological system as a state vector in a high-dimensional feature space. The goal is to compute the optimal perturbation vector (using the LINCS L1000 dataset) required to transition a system from a sub-optimal to a “High-Efficiency” state. This challenge is framed as a Semi-Supervised Recommender System problem involving: (i) Representation Learning (e.g., Autoencoders, Contrastive Learning) to compress transcriptomic data; (ii) Metric Learning to cluster effective perturbations using “anchor” data; and (iii) Zero-Shot Ranking to predict novel candidates.

While the application is in gene therapy, the core challenge is purely computational. No biological knowledge is required, making this project ideal for students with a strong interest in Deep Learning (PyTorch/Keras/TensorFlow) and working with large matrices.

References:

Machine Learning Methods for Population Clustering using Ancient DNA Datasets

Supervisor: Prof Matthieu Defrance

This research project aims to develop novel machine learning methods specifically designed for population clustering using ancient DNA (aDNA) datasets. Paleogenetic data presents unique computational challenges, including high missingness, DNA damage, and low coverage, which often hinder traditional analysis. The core objectives of this work are (i) Method Development: creating and validating machine learning approaches optimized for these specific characteristics; and (ii) Comprehensive Benchmarking: systematically comparing classical methods (such as PCA and MDS) with modern approaches across simulated and real-world datasets.

This project is specifically tailored for students in computer science. While it involves genomic data, no prior knowledge in biology or genetics is required—only a willingness to understand basic biological concepts necessary for the implementation.

Genomic Resources & References:

Methods for omics data clustering

Supervisor: Matthieu Defrance

Clustering analysis is routinely performed on omics data to explore or discover underlying cell identities. The high dimensionality and significant sparsity of these data (with false zero count observations) make clustering computationally challenging. This project studies state-of-the-art techniques for omics data clustering, emphasizing neural network approaches for initial data embedding.

Reference: https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-021-04210-8

Contact: Matthieu Defrance (matthieu.defrance@ulb.be)

⛔️NO LONGER AVAILABLE⛔️Fast variable selection without shrinkage

Supervisor: Maarten Jansen

Selecting optimal models from broad non-nested model spectrums can be driven by criteria balancing good training set prediction and model complexity. Optimization over variable numbers is combinatorially complex and not feasible for high-dimensional data. This problem can be approximated by replacing the counting measure with a sum of magnitude estimators, changing a combinatorial problem into a convex, quadratic programming one.

This thesis applies variable selection in sparse inverse problems, or in deblurring and denoising images, using gradient projection or iterative thresholding.
(NO LONGER AVAILABLE)