Stavanger AI Lab

Artificial intelligence and cerebral stroke

Associate Professor Mahdieh Khanmohammadi has been working on a project involving image analysis of acute stroke. The goal has been to develop an AI tool that can help doctors make better and faster decisions when a stroke is suspected. PhD students Luca Tomasetti and Liv Jorunn Høllesli worked on the so-called twin project until spring 2024.

Luca Tomasetti's part of the twin project mainly involved developing new automatic methods for image diagnostics using machine learning. He used images from CTP scans (computed tomography perfusion) as input for an artificial intelligence network that can segment areas in the brain with reduced blood supply. In other words, to identify the areas in the brain that should be treated for stroke.

Liv Jorunn Høllesli analyzed the relationship between damaged tissue at patient admission and after treatment. The project continues with a new PhD student at UiS, Sazidur Rahman. He will continue working on the machine learning part of the project.

Detects cerebral stroke using artificial intelligence

Khanmohammadi has recently been contacted by a group of researchers from Aberystwyth university in Wales, UK and research partners in Istanbul, Turkey who are interested in collaboration on image analysis of stroke. They collect patient data after the stroke has occurred, for the rehabilitation of the disease itself. So far, Khanmohammadi's research has focused on what can be done to simplify the diagnosis of a stroke. Now they are ready for phase two – namely artificial intelligence to assist in the treatment of stroke. She recently visited Istanbul to meet with the collaborators and visit the clinic where they collect data.

Another ongoing research project concerns Parkinson's disease. Here, researchers use artificial intelligence to study the progression of the disease. The goal is to develop good tools for neurologists. However, there are challenges in implementing AI tools in healthcare due to costs. Mahdieh Khanmohammadi emphasizes that they are still in the development phase and that the long-term goal is clinical implementation.

What does user simulation mean for generative artificial intelligence?

The article "User Simulation in the Era of Generative AI" by Krisztian Balog and ChengXiang Zhai provides an overview of user simulation and its significance for generative artificial intelligence.

User simulation involves mimicking how humans behave when interacting with information systems such as search engines and recommendation systems. These are the main points of the article:

• User simulation involves creating an intelligent software agent that mimics the actions of a human user interacting with artificial intelligence (AI). This is important for modeling and analyzing user behavior, generating synthetic data for training, and evaluating interactive AI systems in a controlled and reproducible manner.
• User simulation draws on concepts from psychology, economics, and human-computer interaction to create accurate models of user behavior. Large language models (LLMs) have made it possible to simulate complex user actions and have been used in various simulation tasks.
• User simulation has several applications:

1. User modeling: Creating realistic simulations of how users interact with a system to improve personalization.
2. Data augmentation: Generating synthetic user interactions to enhance the training of machine learning models.
3. System evaluation: Measuring system performance based on interaction data generated by simulated users.

The article discusses the need for further research across various disciplines to improve user simulation technology. This includes developing more realistic and varied user simulators, as well as integrating advanced machine learning techniques.

The development of realistic user simulators is closely linked to the goal of creating intelligent software systems with human-like intelligence, known as Artificial General Intelligence (AGI).

The article emphasizes the importance of collaboration between academia and industry to validate user simulators and promote innovation in this field.

Using machine learning techniques to detect Parkinson's symptoms

Associate Professor Florenc Demrozi is doing a research project that focuses on using machine learning to detect symptoms of Parkinson's disease, particularly motor symptoms. Motor symptoms significantly impact daily life, making it crucial to understand them better. One of the most challenging symptoms is "freezing of gait" (FoG), where a person suddenly stops and cannot move their feet, often triggered by distractions like a narrow hallway at home. These episodes can last from one to several seconds and can lead to falls, which are especially dangerous for vulnerable and elderly patients.

In such research direction, together with national and international partners, they are developing both new machine learning approaches and wearables devices able to detect FoG patterns and intervene on the users. This combination of algorithmic (i.e., machine learning) and electronics (i.e., wearables) has two main functions: predicting FoG two to three seconds before it happens and helping the person get out of the frozen state through stimulation. The goal is to automate this mechanism, but due to the subjective experience of FoG, it is challenging to create a universal system for all patients.

Currently, the researchers are collecting data; however, due to the high variability of FoG and motor symptoms among individuals with Parkinson's disease, obtaining sufficient and diverse datasets is challenging. While research in this field is ongoing, only a small number of studies provide open-access data, limiting the ability to develop and validate machine learning models effectively. To overcome this challenge, the team is leveraging zero-shot (ZSL) and few-shot (FSL) learning techniques, allowing models to recognize and classify FoG episodes even with minimal available training data. Both the software (AI models) and hardware (wearable devices) are being developed in-house by the research team. The next step is to establish an open collaboration with a hospital to test the AI system on new Parkinson’s patients and evaluate its effectiveness in detecting and predicting FoG episodes in real-world scenarios.