FLAMES annual meeting 2024 @UAntwerpen:

The use of AI and state-of-the-art methodology in research and education


The Flames Annual Meeting is an annual one-day symposium targeting hot topics of artificial intelligence (AI) and the place of statistics and methodology therein, at which prominent speakers will share their insights. This year, the FAM is hosted by UAntwerpen and will take place on Friday the 19th of April, 2024.The topic of this symposium will be “The use of AI and state-of-art methodology to research and education”.

Using AI daily but also understanding the danger of using it blindly helps to develop new knowledge and perspectives. However, there are many advantages in using AI but also challenges which we will tackle in our keynote lectures. Young researchers are also invited to share their work and experiences of involving AI in their research and education.

The Flames Annual meeting will take place at City Campus of University Antwerpen, Hof van Liere building, Prinsstraat 13, 2000 Antwerp. The FLAMES annual meeting will be hold in person. If you have any questions regarding this year's annual meeting, feel free to contact us via statua@uantwerpen.be.

Provisional schedule:

Parallel Sessions

Session 1:

  • Chat GPT - More than a personal assistant: harnessing AI in qualitative research
    Dr. Dries Van Gasse (UA)
  • Cross-sectional associations between 24-hour movement behaviors and cardiometabolic health among adults with chronic diseases: insights in compositional data analysis and time reallocations
    Iris Willems (UGent)
  • Architecture of knowledge as AI precursor
    Olha Sobetska (VUB)
  • Exploring stability of VA-estimates for school accountability systems
    Tom Van Ransbeeck (KULeuven)

Session 2:

  • Optimizing daily commute restrictions for mitigating COVID-19 spread in Belgium: A multi-agent multi-armed bandit approach
    Sebastiaan Weytiens (UHasselt, VUB)
  • Post hoc calibration of medical segmentation models
    Axel-Jan Rousseau (UHasselt)
  • The influence of resolution on the predictive power of spatial heterogeneity measures as biomarkers of liver fibrosis
    Jari Claes (UHasselt)
  • Using AI for better disease management in Alzheimer's disease
    Dr. Lars Costers (VUB, icometrix)

Abstracts of FAM invited speakers

Epidemic policy optimization with reinforcement learning - Prof. Pieter Libin

Epidemics of infectious diseases are an important threat to public health and global economies as we were recently reminded of by the SARS-CoV-2 pandemic. For future epidemic emergencies, it is important to devise optimal mitigation strategies. To properly understand epidemic processes and to study emergency scenarios, epidemiological models are key. Such models enable us to make predictions and to study the effect of prevention strategies in simulation. Even with such models, the development of mitigation strategies which need to fulfill distinct criteria (i.a. prevalence, mortality, morbidity, cost) remains a challenging process. As such, there is a great potential to use reinforcement learning to optimize mitigation policies and support decision makers. In this talk, Prof. Libin will discuss how reinforcement learning can assist decision makers in combination with different types of epidemiological models, addressing both the state of the art and the challenges ahead.

Making AI models for health useful - Prof. Kris Laukens

The progress of AI technology is widely expected to have a strong impact on the future of healthcare. Both algorithmic innovations (in machine learning architecture) and the ever-increasing emergence of rich data generating technologies (from wearables to omics platforms) are driving the development of clinical prediction models. Despite these promising developments, the number of computational models that have actually made it to the clinic and helped the patient is still very low. We will first share some of our experiences in developing clinical prediction models. Second, through a number of practical use cases, we will discuss the factors that contribute to the translation of such models into clinical decision support tools that are useful to the patient. We will end with a set of recommendations for the development of clinically useful AI models.

AI research to detect cell type from high-resolution whole slide images - Prof. Dirk Valkenborg

Digital pathology has revolutionized tumor tissue analysis through Whole Slide Imaging, yet processing high-volume data poses challenges. AI tools are vital for insight extraction followed by spatial statistical analysis to quantify tumor heterogeneity. This biomarker predicts therapy outcomes by cataloging tumor microenvironments, particularly immune cell infiltration crucial for immunotherapy efficacy. High-quality, consistent image capture is pivotal. We investigate a state-of-the-art scan device's properties for whole slide imaging and evaluate nucleus segmentation using deep learning. We analyze cell segmentation performance across various scan settings and resolutions. Interestingly, differences in segmented cells between high- and low-resolution images may hold clinical significance. This research underscores the importance of optimizing scan settings for effective tumor analysis, enhancing digital pathology's clinical utility.

Spatial Statistics in Digital Pathology - Prof. Christel Faes

Tumor microenvironment is the social network of cells (e.g., tumor cells, immune cells, stromal cells, blood vessels, and nerves) that interact with each other and with the cancerous cells. In short, this is often referred to as 'the tumor' but actually represents an intricate ecosystem that gives rise to tumor heterogeneity. The presence of tumor heterogeneity has postulated the primary cause of treatment failure of current strategies. Intra-tumor heterogeneity has namely an impact on tumor biopsy sampling, therapeutic strategy, and therapeutic failure. Using digital pathology has the unique advantage of precisely locating normal and cancerous cells, allowing to study the tumor microenvironment using quantitative statistical methods. Thus, instead of a visual inspection of the spatial heterogeneity by the pathologist, spatial point pattern models that quantify the cell-to-cell interaction can help to investigate tumor heterogeneity and the interaction between the various cell types. A review of Nawaz and Yuan on spatial heterogeneity in cancers outlined the importance and relevance of spatial statistics in describing the tumor microenvironment. In this presentation, an overview of quantitative methods to characterize the spatial distribution of cells in the tumor tissue and the interaction amongst different types of cells is provided, inspired by recent developments in the field of spatial statistics to describe the biodiversity of ecosystems.

Using quantitative and qualitative methods to analyze and optimize 24-hour movement behaviors - Prof. Marieke De Craemer

Interventions targeting lifestyle behaviors are critical for improving health, especially in clinical populations. In the past, lifestyle interventions predominantly focused on behaviors in isolation (e.g., physical activity, sedentary behavior, sleep). However, all the behaviors that people conduct within a 24-hour day can be categorized as either physical activity, sedentary behavior, or sleep. Within 24 hours, a change in time spent on one or more behaviors (e.g., an increase in physical activity) affects the duration of at least one of the other behaviors (e.g., a decrease in sedentary behavior and/or sleep). Therefore, these behaviors are co-dependent and interchangeable. Recent research now investigates physical activity, sedentary behavior, and sleep as a whole, the so-called 24-hour movement behaviors. Due to the co-dependency of 24-hour movement behaviors, standard data processing methods are inadequate. Compositional data analysis is an innovative data processing method to appropriately account for the co-dependence and finite nature of a 24-hour day in which 24-hour movement behaviors are performed. Studies show a less optimal 24-hour composition in several age groups and populations with low volumes of physical activity, high volumes of sedentary behavior, and too short or too long sleep periods. Therefore, it is important to develop interventions focusing on 24-hour movement behaviors to improve people’s 24-hour composition. However, the long-term adherence of current existing lifestyle interventions is poor, with high attrition rates and non-adherence. Most lifestyle interventions are developed using a top-down approach, in which researchers develop interventions and impose participants to implement them. A participatory approach (e.g., co-creation) is a bottom-up, high potential solution to the low adherence rates since it is known to increase the autonomous motivation, empowerment, and ownership of the target group as they are directly involved in the development and decision making of the intervention. In the presentation, both quantitative (i.e., compositional data analysis) and qualitative (i.e., participatory research) will be discussed in general and with concrete examples from the MOVEUP24 research group.

STATS – Sensing & Testing Analytics Toolbox for Soldiers - Prof. Steven Verstockt

Military personnel belong to the group of so-called “high performers”; they must constantly perform physically and mentally and are trained to do so. The physical and mental health and fitness of military staff is paramount to the efficiency of their organization’s core operations. Attrition of recruits, depression or burnout, inadequate fitness and food intake, fatigue, injuries, and illnesses can challenge and damage that efficiency. STATS therefore focuses on early detection of suboptimal performance and dysfunction at the physical, mental, or health level. Detection is based on objectively measured sensor data and specific military performance tests. More info on STATS can be found at: https://www.victoris.be/stats-sensing-testing-analytics-toolbox-for-soldiers/

Challenges of AI in education - Prof. Dimitri Mortelmans

[No abstract provided]

Architecture of Knowledge as AI precursor - Olha Sobetska (VUB)

The purpose of this talk is to provide a historical perspective on AI, highlighting how human thought has shaped its modern concept throughout history. Given the persisting "black box problem" in AI, it's helpful to present some products of human thought which can offer insight into the essence of common knowledge organization and some AI algorithms, such as “Ars generalis” of Ramon Llull, "Law of the good neighbour" by Aby Warburg, and the Project "Mundaneum" by Paul Otlet and Henri La Fontaine. This principle of analogy can be a valuable educational tool, demonstrating the logic of human thought beyond just theory and formulas.

Cross-sectional associations between 24-hour movement behaviors and cardiometabolic health among adults with chronic diseases: insights in compositional data analysis and time reallocations - Iris Willems (Ugent)

Adults with chronic diseases (i.e., type 2 diabetes mellitus (T2DM), overweight, obesity) have a higher risk of health complications and mortality. Evidence suggests that 24-hour movement behaviors (24 h-MBs; i.e., physical activity, sedentary behavior, and sleep) play a crucial role in managing cardiometabolic health. Therefore, it is of interest to explore the 24 h-MBs among adults with chronic diseases and how 24 h-MBs are associated with their cardiometabolic health. 24 h-MBs were collected via an accelerometer (Actigraph wGT3x+BT) in combination with a diary. Accelerometer data were analyzed using raw data processing techniques, i.e., the R package GGIR. The following cardiometabolic variables were considered: Body Mass Index (BMI), waist circumference (WC), HbA1c, fasting glucose, triglycerides, total-cholesterol, HDL-cholesterol, LDL-cholesterol, blood pressure. Regression models using compositional data analysis explored associations with cardiometabolic variables. If the overall composition (ILR coordinates) was significantly associated with a cardiometabolic health outcome, theoretical time reallocation model predictions were conducted to quantify the associations in meaningful terms. The study revealed variations in time estimates of 24 h-MBs based on the data processing methods used. Among adults with T2DM, differences in 24h-MBs were observed across various weight categories. Specifically, the 24h composition of adults with T2DM being obese consisted of less sleep, light physical activity (LPA), and moderate to vigorous physical activity (MVPA) and more sedentary time (ST) per day as compared to adults with T2DM being overweight or normal weight. Regardless of weight category, the largest associations were found when reallocating 20 minutes a day from ST into MVPA for BMI and HDL-cholesterol as well as from ST into LPA for triglycerides. Moreover, these associations were different when stratifying people by short-to-average (7.7 h/night) versus long sleep (9.3 h/night) period. This study highlights the importance of 24 h-MBs in the cardiometabolic health among adults. Shifting time from ST and/or sleep toward LPA or MVPA might theoretically benefit cardiometabolic health.

Optimizing Daily Commute Restrictions for Mitigating COVID-19 Spread in Belgium: A Multi-Agent Multi-Armed Bandit Approach - Sebastiaan Weytjens

Pandemics have a significant impact on the well-being, health, and economy of a society. While various preventive measures exist (i.a., restricting daily commute), deciding on strategies that lead to their most effective and efficient use remains challenging. To this end, compartmental metapopulation models (i.e., the population is subdivided into geographically separated subpopulations) are essential to assist decision-makers in determining the best strategy to curb epidemic spread across different regions. As we have seen during the COVID-19 pandemic, people would get infected at work, for example, and unknowingly spread the disease when returned home, which can be prevented by making coordinated decisions on restricting daily commute. However, simulating such metapopulation models is computationally intensive, and therefore it is pivotal to identify the optimal strategy using a minimal number of model evaluations. Consequently, we explored the use of Multi-Agent Upper Confidence Exploration (MAUCE), a multi-agent multi-armed bandit approach that focuses on cooperatively optimizing a common objective (e.g., hospitalizations). MAUCE leverages loose couplings in the coordination graph to achieve efficient exploration and exploitation in multi-agent problems, which aligns well with the commute network between regions where the optimal policy for each region depends on joint decisions. We will discuss the MAUCE framework and its application to the COVID-19 compartmental metapopulation model, along with initial results demonstrating its potential to identify optimal restriction policies in an epidemic metapopulation setting.

Chat GPT - More than a Personal Assistant: Harnessing AI in Qualitative Research - Dr. Dries Van Gasse

In the rapidly evolving digital age, artificial intelligence (AI) has permeated various facets of academia, with qualitative research being no exception. This presentation explores the transformative potential of AI in enhancing qualitative research across the different stages of doing qualitative review. Chat GPT, often perceived as merely a personal assistant, transcends this role by offering a myriad of applications in academic research. Its natural language processing capabilities enable it to understand, generate, and manipulate human language, making it a powerful tool in reviewing literature, preparing data collection, the early classification of research materials, and its applicability in reports. Using AI brings the potential to increase not only the efficiency but also enhances the quality of research by minimizing human error and bias. However, it is crucial to understand what the technology can and cannot. In this talk, we will discuss what artificial intelligence is and how it can help us in our work.

Using AI for better disease management in Alzheimer's disease - Lars Costers (VUB)

Lars will present the results of a study that evaluates an AI-based software tool designed to assist radiologists in interpreting brain MRI scans for Amyloid-related imaging abnormalities (ARIA) in Alzheimer's disease patients. ARIA monitoring is crucial for guiding treatment decisions with amyloid-β–directed monoclonal antibody therapies. Sixteen radiologists participated, analyzing 199 retrospective cases both with and without the software. The findings indicate that the software significantly enhances the accuracy of ARIA detection, making it a valuable asset in the clinical management and safety monitoring of Alzheimer's disease patients under treatment.

Post hoc Calibration of Medical Segmentation Models - Axel-Jan Rousseau

Deep Neural networks have become state-of-the-art when it comes to medical image segmentation. However, the calibration of these models is an often overlooked aspect of the model’s performance, even though a well-calibrated model can provide a measure of uncertainty for its predictions that can be beneficial for the user or in other downstream applications. We investigated several post hoc calibration methods on several medical segmentation tasks. We introduced two spatial extensions of Platt scaling and beta calibration that leverage spatial information available in the segmentation map. We find that post hoc calibration methods are easy to implement and fast, and provide a large improvement in the calibration of the model with only a small change in segmentation quality.

The influence of resolution on the predictive power of spatial heterogeneity measures as biomarkers of liver fibrosis - Jari Claes

Spatial heterogeneity of cells in liver biopsies can be used as a biomarker for disease severity of patients. This heterogeneity can be quantified by non-parametric statistics of point pattern data which make use of an aggregation of the point locations. The method and scale of aggregation are usually chosen ad hoc, despite values of the aforementioned statistics being heavily dependent on them. Moreover, in the context of measuring heterogeneity, increasing spatial resolution will not endlessly provide more accuracy. The question then becomes how changes in resolution influence heterogeneity indicators and subsequently how they influence their predictive abilities. In this study, cell level data of liver biopsy tissue taken from chronic Hepatitis B patients is used to analyze this issue. Firstly, Morisita–Horn indices, Shannon indices, and Getis–Ord statistics were evaluated as heterogeneity indicators of different types of cells using multiple resolutions. Secondly, the effect of resolution on the predictive performance of the indices in an ordinal regression model was investigated, as well as their importance in the model. A simulation study was subsequently performed to validate the aforementioned methods. In general, for specific heterogeneity indicators, a downward trend in predictive performance could be observed. While for local measures of heterogeneity, a smaller grid-size is outperforming, global measures have a better performance with medium-sized grids. In addition, the use of both local and global measures of heterogeneity is recommended to improve the predictive performance.


2024 — © Flames All rights reserved

Cancellation policy | Privacy Policy