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 walks through a medical journal article that uses machine learning to highlight the concepts from the previous video about crucial parts of these papers.
Weekly reading
The FDA issued a draft of its guidance for machine-learning-enabled software functions. It applies to software devices “which can learn through data without being explicitly programmed and for which modifications are implemented either automatically by software or manually with human input.”
Background
In 2019, the FDA put out a framework that introduced the concept of a predetermined change control plan (PCCP). A PCCP is the idea that the software can make changes on its own without going through a full review process as long as it doesn’t impact the device’s safety or effectiveness. When the device is submitted, the manufacturer can basically say “we plan to update in X way for specific populations or to improve safety” and be able to skip approval when they actually make those changes.
Terminology
The FDA uses “tuning” instead of “validation” as the ML community would understand it, because the FDA has a specific definition of validation related to proving that the device itself can fulfill its intended use.
Requirements
Description of Modifications
- Will the modifications be done automatically or with human feedback?
- Will the modifications be done on all devices or only within certain subsets?
Modification Protocol
- How will the data be managed/stored/annotated/curated/stored?
- How will the data be retrained?
- How will the performance be evaluated?
- How will software updates be performed and communicated to users?
Impact Assessment
- Compare the current version of the device with the device without modifications
- Assess risks/benefits of each modification
- Describe the impact of each modification on another modification, and how the total impact of all modifications
- Describe the impact of modifications to the functionality of the device
Fun Facts
From Invention and Innovation: A Brief History of Hype and Failure
- Inventions that turned from great to terrible: leaded gasoline, DDT, chlorofluorocarbons
- Inventions that were supposed to change the world and didn’t: airships, nuclear fission, supersonic flight
- Inventions that we keep waiting for: Travel in a hyperloop, nitrogen-fixing grains, controlled nuclear fusion
This Week’s Top Stories
- I strongly believe physicians will need to learn to “speak” the language of AI, or we risk being either left behind or using products we don’t understand that aren’t optimized for patients or physicians. This article describes how the workforce of the future will need to be “bilingual” with AI as a primary language.
- The Italian data-protection authority blocked ChatGPT. They don’t like that the system is collecting personal data for algorithm training.
- The Coalition for Health AI released a framework for trustworthy AI in healthcare. They focus on usefulness, safety, privacy/security, fairness, and transparency.
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