Skip to content

Our research

This group actively conducts research in several topics. The themes Research Themes describes the high-level research trajectory, and are directly related to the previously published articles and to ongoing and forthcoming projects.

Since we have many projects ongoing at the same time, not all members of the group are part of every project, and we invite and collaborate with different researches outside the group in projects we deem to align with our purpose as a group, to facilitate a broad and well-established foundation for policymakers and healthcare units to rely on.

Published research

ChatGPT (GPT-4) versus doctors on complex cases of the Swedish family medicine specialist examination: an observational comparative study

Theme: Theme Two - Evaluation of AI-models

Collaborators: This was an internal, PETRA-group project.

Recent advancements in artificial intelligence, such as GPT-4, have shown promise in answering medical multiple-choice questions. However, how well can AI handle free-text assessments for complex primary care cases? This study compared GPT-4's responses to those of real doctors using anonymized cases from the Swedish family medicine specialist examination. While GPT-4 achieved a mean score of 4.5 on a 10-point scale, it fell short of both randomly selected doctor responses (6.0) and top-tier doctor responses (7.2). This underscores the current limitations of AI in nuanced medical decision-making. Read the full article here.

The percentage of the maximum score for each subject category achieved by each groupThe percentage of the maximum score for each subject category achieved by each group. Statistically significant differences (p<0.05) compared with group A, the random doctor responses, are marked by an asterisk (*).

Towards evidence-based practice 2.0: leveraging artificial intelligence in healthcare

Theme: Theme Five

Collaborators: The study originated from Halmstad University, and consisted of a collaboration between Halmstad Univeristy, Linköping University and the PETRA-group.

This review examines how AI can evolve evidence-based practice (EBP) into "EBP 2.0" by addressing current challenges in its three pillars: evidence, clinical experience, and patient preferences. Traditional EBP struggles with slow evidence generation and dissemination, the difficulty of applying general research to individual patients, and biases in clinical judgment. Patient involvement in decision-making also faces obstacles.

AI offers solutions by accelerating research, improving systematic reviews, and enhancing clinical trials. It can possibly augment clinical outcomes by processing vast data for more accurate diagnoses, particularly in image analysis, and provide innovative simulation training. AI also has the potential to free up clinician time, potentially leading to better patient interactions and increased patient autonomy.

However, concerns exist regarding AI's potential to perpetuate biases, the "deskilling" of clinicians due to automation, and privacy risks with sensitive health data. The article concludes that while AI holds significant promise for EBP, further empirical research is essential to understand its full impact and ensure its responsible integration into healthcare.

Artificial Narrative Medicine

Theme: Theme Five

Collaborators: This was a collaboration between the PETRA-group member Sundemo and the Uppsala University and Harvard University affiliated researcher Charlotte Blease.

This publication outlines a new concept, Artificial Narrative Medicine. It explores the role of AI in medicine, focusing on its impact on patient narratives and the doctor's role. AI shows impressive capabilities in diagnostics and empathy, raising questions about its potential to fully replace medical tasks, supported by the philosophical concept of functionalism.

The article highlights the critical role of stories in narrative medicine for diagnosis and treatment. It presents data showing AI models like Med-Gemini achieving high accuracy in medical exams (91% on the USMLE by March 2024) and even outperforming human doctors in medical quality and empathetic responses in a study.

Despite AI's performance, the article questions whether AI truly "listens," pondering what is lost when the human element is removed. While AI can automate many medical tasks, the value some patients place on being heard by a conscious human listener remains a significant consideration. The authors suggest a selective approach to artificial narrative medicine, where AI handles routine cases, allowing human doctors to focus on narratives with deeper existential and psychological weight.

Ongoing research

  • The Survey study, "AI Entering Healthcare: What Requirements Should We Set?", conducts a survey among clinicians and the general public to better understand how accurate artificial intelligence tools must be for them to be used in a healthcare setting, as well as who bears the responsibility for errors that occurs as a result from the use of artificial intelligence in healthcare.
  • The Perspective Methasynthesis project aims to synthesize current qualitative research on the perspectives of clinicians and patients on artificial intelligence in primary care settings.
  • The Question study, "Can AI Address Primary Care's Knowledge Requirements? A Study of Clinically Relevant Questions", aims to categorize and quantify the types of questions that doctors and nurses in Swedish primary care might ask a medical chatbot. The collected questions will inform future studies, including benchmarking AI performance in answering medical queries, comparing AI-generated responses to human expertise, and creating datasets for training and evaluating medical chatbots in a Swedish healthcare context.
  • The Qualitative study, "Perceptions and Attitudes Toward AI as a Tool in Primary Care", explores perceptions and attitudes toward the use of AI, particularly large language models, in Swedish primary care as clinical decision support tools for healthcare professionals or direct interfaces for patients. Through qualitative interviews, it aims to identify key themes and patterns to guide the implementation and development of AI in healthcare.
  • The Synthetic EHR study, "Creating an Open, Synthetic Medical Record Dataset in Swedish for AI Development in Healthcare", aims to create an open dataset of synthetic medical records in Swedish, mimicking the language and structure of real patient records while ensuring complete anonymization. The dataset will support AI development in healthcare by providing realistic training data for language processing tasks.

The Primary care Emergent Technology Research and Advancement Group