🤝 How a leading hospital uses AI to learn from medical errors #2

Today’s Insights:

• How a leading hospital uses AI to learn from medical errors

• The world’s smallest walking robot: microscale measurements

• ChatGPT-4 outperforms others in medical AI model comparison

• News from world’s largest radiology conference Radiological Society of North America (RSNA)

• Bonus: Google is helping everyone build AI for healthcare applications with open foundation models

 TODAY IN AI

• How a leading hospital uses AI to learn from medical errors: Karolinska University Hospital in Sweden, known as one of the most transparent hospitals in the world, uses an AI technology called CRAB to analyze patient outcomes and identify areas for improvement. The hospital achieves high survival rates and very low complication levels. The system helps doctors learn from complications rather than assign blame, helping to distinguish between complications expected based on patients' conditions and unexpected ones that may be preventable.
  
  "We are trying to use this as a way of finding where we are bad. We want to find all the things that really are sub-optimal, and see if we can do something about it." — says Gunnar Sandersjöö, head of the hospital's trauma centre.

• The world’s smallest walking robot: microscale measurements: Cornell researchers in physics and engineering have developed the smallest walking robot to date, designed to interact with visible light waves and move independently. This allows it to navigate precise locations, such as within tissue samples, to capture images and measure forces at the microscopic scale of the body's smallest structures.
  
  “These robots are 5 microns to 2 microns. They’re tiny. And we can get them to do whatever we want by controlling the magnetic fields driving their motions.” — Itai Cohen, professor of physics and a co-author of the study.
  

• ChatGPT-4 outperforms others in medical AI model comparison: MedConceptsQA is a new open-source tool for testing LLMs' medical knowledge and reasoning abilities. When tested on various models, including clinically-trained ones, with different level difficulty, most performed poorly with random guessing. GPT-4 stood out, showing 27-37% better accuracy. This benchmark helps evaluate AI systems' potential for healthcare applications and guides future development.

NEWS FROM RSNA 2024

The world’s largest radiology conference Radiological Society of North America (RSNA) annual meeting is taking place currently in Chicago (1-5th December) gathering tens of thousands of imaging, IT and informatics professionals from more than 120 countries, to explore latest software and new clinical innovations.
Here are some news from RSNA:

• Radiology AI breakthrough: Aidoc has introduced its CARE1™ (Clinical AI Reasoning Engine, Version 1), a groundbreaking AI model trained on millions of CT exams. This model aims to significantly reduce diagnostic delays in clinical practice by enhancing precision and speed in radiology. The first application based on CARE1 has been submitted for FDA review, marking a pivotal step in integrating advanced AI solutions into clinical workflows.

• Faster cancer diagnosis: Lunit, in collaboration with AstraZeneca, has developed an AI tool that predicts genetic markers for non-small cell lung cancer (NSCLC) directly from standard hematoxylin and eosin (H&E) stained slides. This innovation eliminates the need for molecular testing, thereby accelerating the cancer diagnosis process.

If you want to get more news from RSNA conference, read article from Healthcare IT News and RSNA.

BONUS

Google is helping everyone build AI for healthcare applications — they launched open source foundation model: Health AI Developer Foundations (HAI-DEF), a resource to help developers build & implement AI models for healthcare more efficiently.