Faculty of Mathematics, Physics
and Informatics
Comenius University Bratislava

Computer vision seminar - Gabriela Czanner (29.3.2023)

Wednesday 29.3.2023 at 13:10, Lecture room I/9

22. 03. 2023 16.19 hod.
By: Zuzana Černeková

doc. Gabriela Czanner (School of Computer Science and Mathematics at Liverpool John Moores University):
Statistical learning from complex data for better understanding and decision making in neurobiology and medicine

In domains of medicine and neuroscience, there is an increasing number of complex datasets. They have potential to answer questions that have not been answered yet. However, the growth of the new datasets needs to be matched with growth of appropriate flexible statistical learning methods. One of the frameworks that is appropriate is a so-called model-based statistical learning methodology. The key idea is to build a complex model, that captures the complexity of the data, and then do use this model for learning: to do inference about disease or biology or to do a prediction.

Here, we first bring four typical complex data scenarios to illustrate the rich spectrum of problems that complex data can bring. The first scenario is from neuroscience, were we studied spiking behaviour of hippocampal neurons of monkey which was learning a location-scene association task. These high-frequency data (every 1ms) were studied to understand how the neuron was involved in the monkey’s learning. In the second scenario we use data from general practices, linked with hospital records, to find out if there is an increased risk of oesophageal cancer in those who take bisphosphonates to prevent or slow down osteoporosis. In the third scenario, we use longitudinally collected data from epilepsy patients, from several visits, and we aimed to predict as early as possible, which patient is in remission. In the last illustrating scenario, high dimensional spatial data: the images of retina. Our aim was to find out how to learn about risk factors of retinal diseases, as well as how to detect diseases.

We will show how we used statistical models to formulate the problem of data complexity, and how we then used the models for statistical learning: to test statistical hypotheses or to detect disease. In the hippocampus neuron, we developed and implemented algorithm that learns from the fitted model, we showed how we learned about neurons’ role in monkey’s behaviour, we then extended our work to develop novel signal-to-noise ratio for neurons. We will show how we approached the problem of confounding and how found new risk factor for oesophageal cancer. We will share experience of how we developed dynamic procedure for monitoring of epilepsy patients, and how we made it safe. Finally, we will discuss our work on statistical spatial models of retina, how this leads to higher power of hypothesis testing, and to high accuracy of disease detection, comparable to machine learning methods.

We will also show how the illustrating examples led to further research and how these seemingly different scenarios are interlinked. We also discuss future avenues for research to harness complex datasets. 

Doc Gabriela is a Reader in Statistics for Healthcare at School of Computer Science and Mathematics at Liverpool John Moores University, where she leads e-Health group at PROTECT research centre. Her research focuses on development of AI and statistics toward decision support systems for healthcare as well as for understanding of the biological processes in health and disease. Her projects are funded by EPSRC, Dunhill Medical Trust, MRC, NIHR, GCRF totalling £7.8 million. Her previous research was at University of Liverpool, University of Warwick, University of Oxford, Massachusetts Institute of Technology. She has served in independent committees for UK clinical trials for AI validation, and as well as developed statistics courses for Royal College of Ophthalmology, and Royal Statistical Society. Gabriela holds a visiting researcher position at FIIT STU Bratislava, collaborating with BrAInworks research group.