Title Promoter Affiliations Abstract "Automated VIdeo-based assessment of DYSkinesia in cerebral palsy using markerless pose estimation and machine learning (AVI-DYS)" "Elegast Monbaliu" "Rehabilitation Sciences, Bruges Campus" "Movement disorders in dyskinetic cerebral palsy (DCP) are associated with impaired muscle tone regulation and interfere with intentional movements. To treat movement disorders invasive neuromodulation treatments are increasingly applied within DCP. Effective monitoring is extremely important for the indication, evaluation and dosing of these interventions. Current methods to assess movement disorders in DCP are insufficient and time-consuming. In addition hospital measurements are not representative of the real-word situation, as movement disorders may vary considerably during the day, and increase with emotions or pain. Therefore, the aim of the study is to develop an objective measurement technique that can be used in everyday situations. We suggest using supervised machine learning to automatically classifies dyskinetic movement patterns and assess severity, using data extracted by markerless motion from videos. Markerless motion tracking runs already automatically for common movements of abled-bodied persons. We will 1) re-train an existing human model on 4000 unique videos of 120 children with DCP performing everyday tasks 2) train algorithms that maps features from the videos to clinical scoring 3) assess psychometric measurement properties of the best performing algorithms on data not used in the algorithm development and 4) assess feasibility on home-based videos. The study is a first step to develop an objective and easily applicable movement assessment tool for DCP." "Qualitative Evaluation of Machine Learning Models." "Bart Goethals" "ADReM Data Lab (ADReM)" "A common and recently widely accepted problem in the field of machine learning is the black box nature of many algorithms. In practice, machine learning algorithms are typically being viewed in terms of their inputs and outputs, but without any knowledge of their internal workings. Perhaps the most notorious examples in this context are artificial neural networks and deep learning techniques, but they are certainly not the only techniques that suffer from this problem. Matrix factorisation models for recommendation systems, for example, suffer from the same lack of interpretability. Our research focuses on applying and adapting pattern mining techniques to gain meaningful insights in big data algorithms by analyzing them in terms of both their input and output, also allowing us to compare different algorithms and discover the hidden biases that lead to those differences." "Enhancing data analytics for IoT by enabling semantic enrichment of machine learning tasks" "Filip De Turck" "Department of Information technology" "The recent spread of sensors, actuators and mobile devices, comprising the Internet of Things (IoT), provides ample opportunity to improve our quality of life through data analytics. However, as IoT data is bound by the four Vs of Big Data—volume, variety, velocity, and veracity—deriving meaningful insights becomes challenging. Today, two approaches have been employed side by side. Relying on knowledge graphs (KGs) and logical rules, knowledge-driven approaches are able to derive new high-level insights via deductive inference. By making use of semantic enrichment, they are able to improve data quality and consolidate heterogeneous data sources. Conversely, data-driven approaches process raw data by applying a wide array of machine learning techniques to capture inductive knowledge. In order to leverage the benefits of semantic enrichment to improve performance on machine learning tasks, I propose using KG embeddings as a form of semantic feature generation. However, due to their static, relational nature, current embedding techniques are not easily applicable to streams of sensory data. To this end, I propose an incremental, schema-aware embedding technique, which is updatable in an online fashion and prioritizes sensory data. Because IoT applications are often critical and machine learning approaches are usually opaque with respect to decision-making, this technique is further integrated into an interpretable, end-to-end decision model." "Machine learning Modeling of Time-dependent Patient Trajectories" "Yves Moreau" "Dynamical Systems, Signal Processing and Data Analytics (STADIUS)" "The increasing availability of large-scale medical datasets has fueled the hope of an acceleration toward precision medicine, driven by data. Indeed, the systematic collection of Electronic Health Records (EHR) across several healthcare institutions has resulted in large numbers of patient records even for low-prevalence diseases and allowed uncovering more specific patterns in the evolution of the disease of individual patients. Ultimately, these uncovered patterns can be used to forecast the future medical trajectory of individual patients and support individual treatment recommendations to improve medical outcomes.Nevertheless, the data collected in daily clinical practice present some specific features that raise challenges for typical machine learning architectures. Modeling of clinical data, and patient clinical trajectories in particular, therefore requires dedicated machine learning methods. In this thesis, I focus on two main requirements that clinical machine learning models should fulfill: the ability to handle irregular sampling and the ability to perform causal inference.Clinical data collected over time is indeed usually irregularly sampled, as it is only collected at medical visits, which do not necessarily happen at fixed time intervals. Moreover, not all possible measurements are collected at each visit. While classical machine learning methods for time series do not naturally support irregular sampling, we propose new architectures based on neural ordinary differential equations that operate in continuous time and can therefore elegantly model irregularly-sampled incomplete time series.  An important application of modeling clinical trajectories resides in the ability to recommend treatments to individual patients, based on their clinical history. However, as most available clinical data arise from daily clinical practice rather than strictly curated clinical trials, inferring the effect of a treatment on individual patients is not as straightforward as mere forecasting. Estimating the individual treatment effect is a causal question and thus requires a causal framework. We thus propose three new causal machine learning models for patient trajectories: a longitudinal individual treatment effect prediction model focused on uncertainty quantification, a model to estimate counterfactuals when patients can be divided into clinical subgroups, and a causal discovery algorithm based on convergent cross mapping. We motivate all our contributions from a theoretical perspective and demonstrate their improved performance over previous state-of-the-art approaches through a series of rigorous numerical experiments. Lastly, we perform an in-depth analysis of a specific clinical use case on multiple sclerosis. We develop a machine learning method to predict the risk of disability progression within two years." "Enabling personalized medicine by optimizing disease treatments with hybrid machine learning" "Sofie Van Hoecke" "Department of Information technology, Department of Electronics and information systems" "Dynamic Treatment Regimen (DTR) are adaptive treatment strategies using a sequence of expert decision rules, ideally one per stage of intervention, to individualize treatments for patients. They are an important tool towards enabling personalized medicine. Attempts are already being made to tackle treatment individualization with machine learning using deep reinforcement learning for better accuracy, however, these solutions require immense amounts of data. However, a lot of use cases in medicine do not have access to large amounts of data, as gathering data is extremely costly or simply unavailable. Therefore, I will investigate the combination of expert knowledge with machine learning in a hybrid dynamic treatment regimen system to enable treatment individualization in these cases. To realize this hybrid dynamic treatment regimen 3 research objectives are defined: 1) Improve personalized treatment outcome prediction with a precision medicine-based hybrid machine learning framework; 2) Design methods to include expert knowledge into reinforcement learning to compensate for small data; 3) Make the framework and methods transferable to different use cases to reduce modeling time and data needs in new use cases. This will provide an end-to-end system that predicts treatment outcomes and suggests treatment using machine learning and expert knowledge." "Physics-enhanced machine learning for domain-aware predictive maintenance." "Sofie Van Hoecke" "Department of Electromechanical, Systems and Metal Engineering, Department of Electronics and information systems" "Unexpected breakdowns in manufacturing occur frequently, resulting in periods of downtime and potential revenue losses. Recent research advancements apply machine learning to predict the occurrence of these breakdowns. Based on the collection of historical data describing the circumstances before these malfunctions appear, models are trained to find patterns in the data that predict the remaining life time of a machine. The main drawback of these approaches is however their reliance on (lots of) data to be able to function well for many types of failures and operating conditions. Therefore, in this research proposal, I aim to guide these models to the right solution by also providing them with physical knowledge available about the machine. Engineers often have detailed knowledge about the physical behavior of the machines they design. For example, when a car engine is running, it should not exceed certain temperatures or noise levels. When it does, something is wrong with the engine, meaning that it will fail soon. Instead of having the models to learn these known patterns from historical data, we will provide them with this expert knowledge from the beginning. This enables them to learn much faster and with less data. In addition, we don't want to retrain these programs for every new environment a machine is deployed in (e.g., different weather conditions). The models should be smart enough to extrapolate important warning signs they have learned from other environments." "Machine learning for fraud analytics" "Wouter Verbeke" "Information Systems Engineering Research Group (LIRIS) (main work address Leuven)" "Fraud remains a major challenge for businesses. The Association of Certified Fraud Examiners (ACFE) estimates that a typical organization loses 5% of its revenues due to fraud. Furthermore, fraudsters continuously adapt their techniques in response to fraud detection efforts, creating a need for adaptive fraud detection systems. Given the abundant availability of data, machine learning techniques seem well suited to tackle this problem. Therefore, the goal is to develop a profit-driven fraud detection system that is able to swiftly adapt to changes in fraudulent behavior by continuously learning from the data. Additionally, given the complexity of real-life organizations, this system should be able to incorporate multiple levels of information. Therefore, the focus will particularly be on ensemble meta-learning schemes. These are observed to be amongst the most powerful supervised learning techniques, while being innately adaptive. Moreover, by customizing the objective function, they can be made cost-sensitive. These new techniques will be evaluated empirically with real-life datasets. This way, the project aims to add to the existing literature on profit-driven analytics, cost-sensitive learning, robust statistics and user and entity behavior analytics for fraud detection." "Machine learning platform for better cardiovascular health assessment and risk stratification" "Tatiana Kouznetsova" "Hypertension and Cardiovascular Epidemiology, Public Health and Primary Care, Kulak Kortrijk Campus" "Precise description of cardiovascular health and risk assessment are essential for optimal prevention and treatment strategies, but require a complex integration of many factors. The specific objectives of this project are: (1) Applying machine-learning (ML) algorithms to the available longitudinal, highly standardized, general population data to assess the complexity of clinical and behavioral variables, cardiac imaging sequences and pathophysiologically relevant circulating proteins and metabolites for construction of precise cardiovascular classifiers. These computer models extract maximum information from individual clinical data, the most sensitive biomarkers and diagnostic imaging phenotypes for better cardiovascular risk prediction. (2) The pre-trained ML models are prepared for integration into an open service platform in collaboration with industrial IT partners in healthcare. Special attention will be paid to the implementation of an intuitive graphical user interface and a standardized set of input-predicting variables that enhance the usability of the platform for the medical community." "Forging connections between machine learning techniques and strongly correlated physical systems" "Jan Ryckebusch" "Department of Electronics and information systems, Department of Physics and astronomy" "In this research project, we aim at developing novel ways for understanding physical systems consisting of many interacting constituents using techniques developed in the field of machine learning. A common use of machine learning models is to classify pictures of different objects into categories. Likewise, machine learning methods can be exploited to find hidden connections and correlations between the different degrees of freedom in a physical system. Computer resources required to apply standard algorithms to the study of complex physical systems, typically scale with the system size in an exponential way. Overcoming this unfavorable exponential scaling is a challenging issue. We will use neural networks as a model for these physical systems and use the physics of the complex system to learn about the operation of the neural networks. In this process the computer optimizes the different parameters in the neural network such that it models the physical system under study. It is our aim to unravel the various steps that the neural network has taken to model the physics. This will allow us to unravel the underlying principles and to forge connections between machine learning and physics, to the mutual benefit of both branches of sciences." "Using machine learning to model the prognosis of multiple sclerosis patients" "Celine Vens" "Public Health and Primary Care, Kulak Kortrijk Campus, Laboratory for Neuroimmunology" "Multiple sclerosis is a complex disease with a highly heterogeneous disease course. In this project, we will develop machine learning methods to study this disease course using a combination of demographical, clinical, genomic, and radiomic data. This study is done on the patient level (individualized prognosis models), on the group level (patient stratification models), and on the population level (identification of prognostic biomarkers). With this study, we address pressing needs from both the clinicians, who want to start the right treatment for the right patient as soon as possible, as well as from the pharmacological industry, which wants to develop new treatments as cost-efficiently as possible. In the process, we will contribute to the machine learning literature on genetic data analysis, multi-target learning (i.e. combining outcomes of different data types), recurrent event survival analysis, and clustering (i.e. dynamic over time). Specifically, the models developed in this thesis will need to deal with sporadic and irregularly sampled time series, missing values (even in the outcome space), and confounding factors. To maximize the probability of integration in clinical practice, we additionally ensure that the models are explainable, reliable, and trustworthy."