Health

Single-cell technologies have evolved dramatically the last decade, having a high impact on the life science domain. Using this type of techniques it is now possible to study and analyze tissues and cells in extreme detail. Using single-cell technologies new insights are becoming available for example in the diagnosis and treatment of cancer, neurological problems, immune deficiencies or research on stem cells. By analyzing the individual cell characteristics we can define new biomarkers and we can finetune the  different patient groups, depending for example on underlying health issues. 

The challenge is that these types of techniques generate a magnitude of data that needs to be analyzed in an effective and efficient way, in order to be able to gain the necessary knowledge and biophysiological insights. AI-tools and techniques are needed and are being researched to make this possible. 

The paradigm shift towards better and more precise healthcare provision is gaining momentum today and holds important advantages.

The research within the Flanders AI Research Program

As mentioned the single cell techniques generate a magnitude of data. For the efficient and correct interpretation of these data sets, we are exploring AI-techniques on the following types of data

• Single-cell transcriptomics data: setting up a process for automatic preprocessing and quality assessment on the data and development of a new method to predict interaction in between cells
• Single-cell imaging data: automatic segmentation of subcellular structure in 3D electro microscopic images

Results

• In collaboration with UZ Gent a new pipeline has been developed to analyze single-cell data from Covid-19 patients in order to gain insights in the immune response on the disease. These data have been put in the public domain in the context of the international Covid-19 Cell Atlas project. In collaboration with UZ Leuven new machine learning techniques were developed for automatic analyses of immune data of Covid-19 patients. This led to the identification of four types of immune responses, in correlating to clinical parameters.
• A new method has been developed, called NicheNet to learn about the interaction in between cells. This method has been applied to datasets on head- and neck cancer and on to data on liver cells, which led to new insights on the interaction in between immune cells and in between liver cells. 

Contacts

• Contactperson: Yvan Saeys (UGent/VIB)
• Research groups: UGent (IDLab, DAMBI, GAIM, KERMIT), imec, VIB

More information

• The COVID-19 Cell Atlas Project, klik hier
• NicheNet

Multiple Sclerose (MS) is a disease that attacks the central nerve system. People with MS benefit from receiving the correct treatment as fast as possible, but finding the correct treatment at the correct moment for a specific person is still a challenge today. Doctors, patients and policy makers need actionable and reliable tools to assist them in decision making. AI tools are being researched to assist them.

The research within the Flanders AI Research Program

In this research we focus on three major clinical challenges:

• What type of treatment is best for which patient? We aim at developing real decision supporting tools to be used in clinical environments.
• Is the treatment effective? We aim at identifying new biomarkers in order to measure the disease status and progression thereof.
• Can we improve the treatment by scaling up? We develop tools and methodologies in order to support the international MS data-mining research program. 

Results

• The Global Data Sharing Project, under co-ordination of UHasselt, engaged over 40 partners worldwide to study the impact of the Covid-19 disease on the MS patient population. Within the timeframe of a few months, the project succeeded in gathering data from over 80 countries. Based on the gained insights, they updated the Covid-19 related recommendations for chronically ill patients for the MS patient group. This advice became available in 14 different languages in a few days’ notice. The multidisciplinary team involved in the Flanders AI Research Program developed and operated the data infrastructure for this initiative.
• We managed to proof-of-concept of a beyond-state-of-the-art decision support system (DSS) to predict progression of MS. By developing and implementing AI algorithms that cope better with missingness of trajectory data and increase the interpretability of the predictions, we managed to deliver a DSS that performs better compared to what is available in literature until now, specifically when it comes to discrimination, calibration and interpretability. 

Contacts

• Contactperson: Liesbet Peeters, UHasselt
• Research groups: KU Leuven (DTAI, STADIUS, ITEC), UAntwerpen (VisionLab), UGent (IDlab, KERMIT), UHasselt (IMIN)  

More information

• Covid-19 and MS - Global Data Sharing Initiative - MS Data Alliance (MSDA)

Medical decisions for diagnosis, therapy planning, and follow-up rely heavily on 2D and 3D medical imaging data of the patient, such as RX, CT, MRI, PET and US. Because of continuous technological advances in medical imaging equipment, the amount of imaging data to be analyzed in routine clinical practice is ever increasing, putting high pressure on the radiological workflow. Radiological reading of medical images is a complex process that requires extensive prior knowledge and expertise to detect and recognize subtle patterns in the images and discriminate abnormal findings from normal variability. More and more clinical applications also benefit from quantification of specific anatomical or functional parameters of the structures of interest, such as position, size, volume, shape, motion, deformation, density, perfusion, diffusion…. Such quantification often involves precise 3D localization and delineation of anatomical object boundaries in the images, of both normal anatomy as well as pathology, which is tedious and time-consuming and hence not feasible in routine clinical practice if to be performed manually by the expert. There is thus a strong need for reliable, automated computer-aided assessment and quantification of medical imaging data for clinical decision support in Radiology, Oncology, Cardiology, Neurology, Surgery and many other medical disciplines.

The research within the Flanders AI Research Program

In this project, we wish to develop machine learning strategies, implement them into AI-tools and validate their clinical application potential in a number of different, but methodologically related clinical decision support applications involving medical imaging data from different application domains, focusing in particular on applications in Radiology and Radiation Oncology. These applications have variable amounts of available training data, and priors over image structures, providing non-trivial and informative test cases for AI methods. Tasks can be broadly categorized into classification, regression, and image/volume segmentation problems.

Direct application of deep learning strategies that have been successful in related computer vision applications to medical imaging data is challenging because of the nature of the data itself (3D, multi-modality), the limited training data that is typically available (hundreds instead of millions of images), the need for expert input for clinically relevant assessment of performance (some errors are more significant than others), and the fact that ground truth may be unreliable due to observer variability induced by ambiguity in the images (disagreement between experts).   

Results

In 2020 this research has been applied to lung images during the Covid-19 pandemic. CT scans of the lungs deliver interesting information in understanding the disease and in the follow-up of the patients. Deep learning techniques are being used to accurately detect lung lesions (abnormal lung tissue). Researchers from the KU Leuven (PSI) and the VUB (ETRO) have developed and compared methods for fast and automatic analyzes of CT lung images. The results have been used in the iCovid project, a pro-bono project initiated by several hospitals and research organizations. The research results have been incorporated by icometrix NV in their cloud based image analyzes product, called ‘icolung’. Over 800 hospitals in Europe use this service on a daily basis. In a further stage, the iCovid project received funding from the EU H2020 program.

Contacts

• contact persons: Matthew Blashko, Frederik Maes (KU Leuven-PSI), Hichem Sahli (VUB-ETRO)
• Research groups: KU Leuven (PSI), en VUB (ETRO), UHasselt (CENSTAT)

More information

• iCovid project: Lung imaging AI.
• icometirx NV: icolung

Epilepsy is a neurological disease that affects around 65 million people worldwide. Despite the different currently available treatments, still 35% of the patients continue to have seizures for the rest of their lives. The quality of life (QoL) of these patients and their relatives is strongly influenced by the unexpected behaviour of epileptic seizures. One option to improve their QoL is to use an automated seizure warning system, which automatically detects ongoing seizures and alarms the patient’s relatives. 

Another option to improve their QoL is to optimize the treatment itself. Currently, treatment evaluation relies on manual seizure diaries maintained by the patient itself, which have shown to have an accuracy under 50%. An objective and automated seizure diary could lead to better patient treatment due to more reliable information for the clinician.

Both applications rely on the automated detection of epileptic seizures, but today a too low accuracy of these algorithms exist due to a wide variety of seizure types, the patient-dependency of epileptic changes, the lack of accurately annotated data, poor data quality and suboptimal multimodal combination. 

The research within the Flanders AI Research Program

The aim is to develop AI tools that are more accurate in predicting a potentiel next epileptic seizure or an early indication of an increased risk thereof. Using this information patients could have better self-management tools and the neurologist will have a better view on the medical condition of the patient. 

Using retrospective analysis of clinical data sets we develop new algorithms in view of three major challenges:

• Better automated quantification of biomedical signals for seizure detection
• Improved multi-modal classification
• An improved and simplified personalisation of data related to seizures

Contacts 

• Contact person: Maarten De Vos, KU Leuven, STADIUS
• Research groups: KU Leuven (STADIUS, DTAI, ITEC), UGent (IDLab, KERMIT, VIB) en UHasselt (CENSTAT)
• In collaboration with UZ Leuven UZ LEUVEN (Wim Van Paesschen)

More information

• Narayanan, A. M. , Seeuws, N.,  Vandecasteele K., De Vos, M. , Bertrand A. and Van Huffel, S. "Epileptic Seizure Detection in EEG via Fusion of Multi-View Attention-Gated U-net Deep Neural Networks", submitted to the 2020 IEEE Signal Processing in Medicine and Biology Symposium (IEEE SPMB 2020).
• G. Vandewiele, F. Ongenae, F. De Turck, GENDIS: GENetic Discovery of Shapelets, submitted to Evolutionary Computation, Arxiv preprint.

Healthcare costs are increasing worldwide. According to some estimations the cost of healthcare could be 2050 ne more than 20% of the BBP of several countries. Non-optimal use of medical resources is one of the reasons that are often stated as a contributor to this increasing cost. Length-of-Stay is an important parameter for hospital planning and management tools, with respect to hospital admissions for diagnosis, planning of surgery and all related logistics. : of the Length-of-Stay can alleviate the strain on existing resources and can augment the quality of care at the same time by for e.g. reducing waiting time for patients. Optimized patient flow can improve patient satisfaction and decrease costs.

The research within the Flanders AI Research Program

Our aim is to develop AI-based tools for prediction of the Length-of-Stay taking into account the specific condition of the patients and the evolution thereof. Different challenges are investigated:

• Defining patient specific domains for handling the heterogeneity in the patient population.
• Handling heterogeneous patient data taking into account a vast number of patient features and their interactions.
• Dynamic tools updating the patient's LoS prediction in real-time as new laboratory tests, or measurements etc. become available.

Results

By employing recurrent neural networks and transfer learning techniques, we built an efficient  LoS prediction model that takes into account patient specificities while sharing common information across medical units and saving hospital computing resources. This new model provides more granular information to the hospital management by giving the prediction accuracy in each medical unit. This illustrated that, feeding all patient data into a single model as it is usually done either over/under estimates the error for these different and unique patient populations.
The AI research methods used focused on long-short memory neural networks to learn temporal information across patient data and transfer learning to initiate model training. As a next step the results will be benchmarked on the MIMIC-IV dataset.

Contacts

• Contact person: Bart De Moor (KULeuven-  ESAT-STADIUS)
• Research groups KU Leuven (ESAT-STADIUS), UGent (IDLab)
• In collaboration with UZ LEUVEN (Frank Rademakers)

More information

Check the use case in the domain "Government and citizens"