Research Themes

To effectively channel our research efforts and ensure impactful progress, the PETRA group's projects are organized under five core themes. These themes provide a framework for innovation and collaboration, as briefly outlined below.

Theme One: AI Benchmark for Swedish Primary Care

Many existing AI models in healthcare have been tested using non-representative databases, such as the United States Medical Licensing Examination (USMLE). Our aim is to develop a benchmark for Swedish primary care that is truly representative, containing questions and tasks from working clinicians during their daily patient related work. The goal is to create publicly available benchmarks, accessible to developers and researchers worldwide. These benchmarks will be developed in three stages, with increasing levels of complexity:

• Oracle AI
• EHR-embedded active assistant
• EHR-embedded agentic AI

Theme Two: Evaluation of AI Models

This broad theme comprises different ways of evaluating how current AI technology is applied to healthcare. It includes work such as our previously published British Journal of Medicine article. Further, this theme may involve evaluations such as:

• Evaluation of AI performance on standardized exams or clinical simulations
• Evaluation of AI performance on specifically designed benchmarks (including our own)
• Qualitative studies exploring user experience and ethical implications
• Clinical studies assessing real-world impact and safety

Theme Three: Development of Clinical Applications

In the design of benchmarks, value, usability and creation of applications have to be taken into account. There are several opportunities to develop tools to aid clinicians, spurring from the research conducted by the PETRA-group.

Theme Four: Data Creation, Methods, and Tools

Robust technical tools, methodologies, and datasets are crucial to enable progress in AI research within healthcare. Currently, two major projects form part of this theme:

• Development of Synthetic Electronic Health Records
• Development and iteration of Large Language Model (LLM)-based automated feedback and correction algorithms

Theme Five: Medical Philosophy and Stakeholder Preferences

The increasing capabilities of AI raise profound questions relating to the value and future role of human clinicians. This is partly a philosophical question but is perhaps even more significantly a matter of human and societal preference. We explore this multifaceted question from various angles, including:

• Philosophical inquiry into the intrinsic value of human clinicians in an AI-augmented healthcare system.
• Survey studies researching the preferences and concerns of patients regarding AI in healthcare.
• Interview studies evaluating the perspectives of both patients and clinicians on the integration of AI.