🔍 Spotlight #11: ⬛️ Black Box AI, 📜 Guidance to build with AI, 🩺 Google Med-PaLM M, 🗣️ Amazon HealthScribe
AI Health Hub, 31/08/2023
⬛️ In Defense of Black Box AI
Podcast
This podcast was a pleasant surprise and a mind-opener. I’ve attended several workshops that discussed the topic of explainable AI, where it was always given for granted that explainability is what we need. In this episode instead, the need for explainability is challenged, in favor of caring only about performance. The podcast is hosted by Reid Blackman, Ph.D., author of “Ethical Machines”, with guest Kristof Horompoly who is currently leading Responsible AI at JPMorgan Chase.
The two explore the space of explainable AI, tackling limitations, algorithmic methods and implications when it comes to data scientists in the industry. The well-know limitation is oversimplification of the explanations that can lead to distortion of the outputs. It is also time consuming, costly, and it has its environemental impact. Besides considering the computational resources for running methods for explainability, companies should consider testing which explainability methods work best for the developed models and then working towards optimization. Of course, the speakers comment that certain regulated fields will and should keep a level of explainability for customers and experts. Now the general consensus still remains that AI outputs should be explainable, regardless of the field.
📜 Understanding regulations of AI and digital technology in health and social care
Tools
This is a website funded by the NHS AI Lab and is meant for people who develop or plan to use AI or digital technology in health and social care. It brings together regulations, guidance, and resources for digital healthcare technologies, for both developers and adopters. It is a tool that applies only to the UK, but I found it very interesting and perhaps a good inspiration for other countries to implement such support materials for entrepreneurs in the health tech space.
🩺 Google Unveils Multimodal Generative AI Model Med-PaLM M for Healthcare
News
Google is working towards a multimodal generative AI model, Med-PaLM M. What does “multimodal“ mean, and why is it important? The term “multimodal” refers to the ability to handle various types of medical data as inputs - like medical images, genomics, and clinical language - and perform different tasks. This is different from the more common “unimodal“ single-task AI systems, which are designed for specific tasks, like interpreting scans. In healthcare, such multimodal solutions make more sense, as medicine is inherently a multimodal discipline. It is the clinician’s role to interpret different types of data, such as notes and lab tests, to come up with diagnoses and treatments. Preliminary evidence suggests that Med-PaLM M can generalize to new medical tasks and concepts and perform multimodal reasoning without specific training.
If you would like to dig deeper, check out the published article by Google DeepMind and Google Research.
🗣️ Amazon launches generative AI-based clinical documentation service
News
Amazon enters the generative AI for clinical documentation space with HealthScribe. It came up with a solution in 2019 called Transcribe Medical, but unlike this latest solution, it did not auto-populate relevant information into the electronic health record (EHR). Why is this space booming? Clinical documentation represents a burden on clinicians, as compiling EHRs can take them away from time with patients and can lead to burnout. Accuracy rates of the solutions by Amazon and competitors Microsoft and Suki were not shared.
🥳 I hope you enjoyed the findings of this week!
For the new subscribers, welcome to the AI Health Hub! I hope you enjoyed this edition, and feel free to like or leave a comment.
🚧 Delay + Searching for contributors
As I continue reading and learning in the space of AI for healthcare, I have been finding it hard to keep up with our weekly appointment here. This edition has been sitting in my drafts for over a month, sadly.
For context, I worked in the medical device industry for more than 3 years, and have just started a doctoral program to research and develop a software for cancer screening to support clinical decision-making using AI. This means very busy times 👩🏻💻!
But this community is here to stay, so I started to look for someone to onboard to help as a contributor 👀