Machine Learning for MDs Weekly Digest
What’s New in ML for MDs
Welcome to the ML for MDs Newsletter. The mission of ML for MDs is to connect physicians interested in machine learning. This newsletter provides the most relevant news, journal articles, and jobs at the intersection of medicine and machine learning.
Weekly watching
This week’s video is an accessible view of transformers and GPT – how the technology works and what the previous iterations of GPT4 have been.
Weekly reading
Eric Topol and colleagues came out with a new paper about Foundation Models for Generalist Medical Artificial Intelligence. This paper looks ahead to a future using more adaptable and flexible AI in medicine.
- Foundation models are the latest generation of AI models, are trained on huge, diverse sets of data and can be used in many scenarios instead of only being able to do one thing
- Most medical models currently are very narrow, ie a model trained on chest X-rays can detect abnormalities but can’t write a radiology report
- Of the 500 AI models for clinical medicine approved by the FDA, most have only been approved for 1 or 2 narrow tasks
- Three main advances: multimodal architectures, self-supervised learning techniques, and in-context learning
- The authors define GMAI models as having 3 main abilities:
- Adapting to a new task with a plain English request and solving previously unseen problems without retraining
- Accept inputs and produce outputs of different types (images, text, lab results etc)
- Reason through previously unseen tasks and use medically accurate language to explain outputs
- Six potential use cases are listed:
- Automatically draft radiology reports that take the patient’s history into account
- Assist with surgical procedures in real time via detection of abnormalities and structures
- Bedside decision support with free text alerts and recommendations
- Interactive note-taking
- Communication with patients via chatbots
- Generate novel proteins for desired properties
- Challenges of GMAI
- Validation of use cases since they would be performing many tasks
- Verification of correctness
- Social biases
- Privacy
- Getting enough medical data for adequate training
Fun Facts
- 40% of newly filed patents in healthcare have an AI component
- Machine learning shows 95% accuracy in reading lips
- Recommendations powered by machine learning make up 80% of movies watched on Netflix
This Week’s Top Stories
- NVIDIA is developing a software platform and an “industrial grade edge hardware platform” to help develop and test AI-powered devices. They’re partnering with Medtronic to develop a device focused on colonoscopies:
- “The GI Genius module uses advanced AI to highlight the presence of precancerous lesions with a visual marker in real-time – serving as an ever vigilant second observer. It processes images using advanced algorithms that can identify and mark abnormalities consistent with polyps, including small flat polyps that might otherwise go undetected by the human eye.
- Studies have shown that having a second observer can increase polyp detection rates and every 1% increase in adenoma detection rate (ADR) reduces the risk of colorectal cancer by 3%.
- Use of the GI Genius module has demonstrated a 14% absolute increase in ADR compared to colonoscopy alone for both flat (42% increase) and polyploid (36% increase) lesions, thus increasing accuracy and reducing the rise of interval cancers which can occur between colonoscopies.”
- An ML algorithm can predict insomnia from medical records using just demographic, laboratory, physical exam, and lifestyle covariates. Given the effect on sleep on chronic pain and therefore opiate use/misuse, this is so important for patients.
- Article in Forbes from the CEO of the Permanente Group makes a case that the government is moving too slowly in adapting to AI in healthcare, specifically by:
- Requiring in-person appointments harms patients
- Trying to regulate ChatGPT is like trying to regulate the telephone
- Not allowing physicians to practice in any state
Community News
- If you haven’t introduced yourself, please do so under the #intros channel.
Thanks for being a part of this community! As always, please let me know if you have questions/ideas/feedback.
Sarah
Sarah Gebauer, MD