Shahar Avin

What’s happened this week in the healthcare & AI world? Here is the interesting news and research: 𝐍𝐞𝐰𝐬: 🗞 An AI triage service has reduced patient wait times by 73% for a GP practice. 🗞 The Gov will launch a Regulatory Innovation Office to help regulators remove barriers to tech in places like the NHS. 🗞 The 2024 State of AI Report was published yesterday by Air Street Capital, and notes the commercial “vibe-shift” from existential threat to sales & usage - Oliver Patel, AIGP, CIPP/E published a good report summary. 🗞 Imperial will trial AI to support cancer diagnosis in general practices. 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡: 👩‍💻 Using deepfakes in psychotherapy, eg for a guided encounter with a deceased loved one, could (unsurprisingly) need new protections (Eg to stop an over-attachment that could stagnate the acceptance of loss).* *this paper is a older than this week but is circulating right now. 👩‍💻An analysis finds that AI for global health is a hyped topic, but in practice the determinants of health are being side stepped. 👩‍💻 Interviews with German med students on AI-CDSS found future clinicians have two different attitudes: “the one relying [on]” vs “the one controlling” the AI. P.S. What item was most interesting to you this week?

30

11 Comments

This is how debates in science should be done.

7

All you need is an explanation? 🗣 ❓   I enjoyed the latest joint RSM Digital Health Section and @Dev&Doc podcast featuring Dr Annabelle Painter, Dr Josh Au Yeung, Zeljko K. on #XAI.   𝐓𝐡𝐢𝐧𝐠𝐬 𝐦𝐞𝐧𝐭𝐢𝐨𝐧𝐞𝐝 𝐭𝐡𝐚𝐭 𝐈 𝐟𝐨𝐮𝐧𝐝 𝐢𝐧𝐭𝐞𝐫𝐞𝐬𝐭𝐢𝐧𝐠: 🧠 We need to trust a solution to accept it's decisions, and part of trust is understanding why - therein lies why we find explainability so important but actually it really boils down to our ability/willingness to trust?   🧠 We frequently use the terms ‘#explainability’ and ‘#interpretability’ with respect to AI as interchangeable, but they mean 𝘥𝘪𝘧𝘧𝘦𝘳𝘦𝘯𝘵 things (see: https://lnkd.in/eKuNMbXx.)   🧠 When we talk about explainable AI – do we mean explainable for each decision or how the system as a whole is working or both? And are they equally important? Also, could we even understand it or find it relevant (e.g. complex mathematical reasoning).   🧠 Whilst we expect AI to be explainable, is this a double standard when we don’t even know sometimes how humans make decisions? 🤷‍♀️ When referring for specialists/tests sometimes we have ☑ that says ‘clinical concern’ for when we can’t explain why a patent needs help – so if we can’t always explain our own intuition and make allowances for that, maybe we need to for AI? 🤔   💡 Fascinating point re: whether having an explanation makes doctors more likely to accept a false result. So could this actually worsen our over-reliance to AI? 😨  Something I saw raised in a Nature letter here published on explainable AI in dermatology. Interesting debate - article: https://lnkd.in/eG8AW_DS  Letter: https://lnkd.in/eYMCuTgv, Response: https://lnkd.in/ebZX_hhS   👩 I agree that sometimes we do anthropomorphise AI and treat it like a human than a tool and that expectation could be dangerous. Lauren Oakden-Rayner gave a lecture several years back and discussed a 2015 study about how pigeons 🐦 could detect breast cancer on pathology images (https://lnkd.in/eD3beHN2 😂 ) and argued how they could be seen as a kind of ‘model’ outputting binary classification with high accuracy but we’d never dream of thinking of a 🐦 = 👩, yet we do this comparison with AI.   🤖 The topic of human-AI interaction and decision making cropped up a few times as an area of much needed research – although I’m certain there must be a lot out there given all the courses/degrees/jobs related to #HCI and #UX? No?   Anyway - lovely episode! Would enjoy hearing from anyone working on exciting projects that cross behavioural & social science / medtech. Always a fan of Freakonomics and how predictably unpredictable we can be with decision making…!

10

3 Comments

The performance of Llama 3 degrades more than Llama 2 when quantized. The current guess is because Llama 3 has been trained on 15 trillion tokens (the longest of any LLM), even the smallest decimal point of a parameter “has a purpose”. I think this also points to another important idea: almost all of the LLM, and possibly other foundation models are undertrained. We are yet to maximize the potential of many large models. Source: https://lnkd.in/grZ_q64k https://lnkd.in/g9PEM3U7

14

Pay attention: change is accelerating. And most businesses aren’t ready for it! In a week where headlines, rightly have focussed on seismic changes to global governance: Syria’s regime change, Ukraine, S. Korea, Sudanese crisis… pay attention to what is happening in business. The hype-cycle of AI continues, with the latest demonstration of text-to-video capability from OpenAI’s Sora model. The personalisation of customer experiences _will_ make and break businesses. Google has announced Willow: the next step towards generalised quantum computing. This _will_ destroy Internet security as we know it, and enable profound new computational capability to unlock new insights. (Link below) We are just missing a breakthrough on nuclear fusion (limitless, clean, carbon-free power)… though that gets easier with AI and Quantum Compute to design & control the reactor. And perhaps a similar breakthrough in the business of space: humans mining asteroids and becoming a multi-planetary species. Though this is coming. Meanwhile: climate change, geopolitical tectonics, educational transformation (personalised AI-powered)… in a world where population, education, connectivity and access to capital are all increasing. My New Year’s resolution… to finally write a book! “Leadership in a world of accelerating change” is my initial working title. Have a cool Yule, and a very merry New Year!

7

2 Comments

This paper intuitively makes much sense. Even if we think about how self-supervision works and going back to methods in digital pathology where we would leverage tiling on whole-slide images as a mechanism, these ideas on Structure Invariant Transformation make sense if we just apply them to adversarial transferability. Two key highlights include: 1. Diverse transformed images are crucial for enhancing the transferability of adversarial examples. 2. Local image transformations can create more diversity while preserving the essential structure of the image. The authors demonstrate through experiments on the ImageNet dataset that SIT, when integrated into the momentum iterative fast gradient sign method (MI-FGSM) attack, achieves significantly better transferability than existing state-of-the-art attacks on both CNN-based and Transformer-based models! This suggests that SIT is a general and effective approach for boosting the transferability of adversarial examples and emphasizes the need to include more robust testing defense mechanisms. I might need to try some fun experiments myself with Remyx AI. 🤔 🗞 Paper: https://lnkd.in/gkRgKvuX 💻 GH: https://lnkd.in/gBD7ruVf #genAI #artificialintelligence #adversarialattack #aisafety #multimodal

11

1 Comment

📊 We are pleased to highlight Adam Steventon (https://lnkd.in/eXVs-4t6), Director of Data Platforms at Our Future Health UK, as our final keynote speaker at MultimodalAI'24. With over 15 years of experience in transforming healthcare through data, Adam's work is instrumental in the UK’s largest health cohort programme. 💡 In his keynote "An incredibly detailed picture of human health: the exciting potential of Our Future Health to prevent, detect and treat diseases," Adam will discuss the construction of a multimodal dataset of 5 million individuals. This dataset aims to represent the UK’s diverse population and enhance our capabilities in predicting, detecting, and treating major diseases through AI. 👨‍🔬 Adam will delve into the roles of artificial intelligence in this ambitious project, highlighting both the opportunities and challenges it presents. His insights will demonstrate the transformative potential of AI and large-scale data in shaping the future of healthcare. 🌐 Join us to learn how data and AI can revolutionise healthcare. For more details, including the full abstract and programme, visit our workshop page at https://lnkd.in/e4jMhG-v. 📅 Don’t miss this opportunity to gain valuable insights from an expert at the forefront of healthcare data transformation. Register by June 10th to be part of this enlightening discussion: https://lnkd.in/ds9awckQ #MultimodalAI24 #AI #Healthcare #DataScience #DiseasePrevention #Innovation

32

1 Comment

Great paper focused on Enhancing #ExternalControlArm Analyses through Data Calibration and Hybrid Designs. It proposes two options when thinking about using #realworlddata to an external control arm for a single arm clinical trial. Read the Full article below: https://lnkd.in/eJK2nqdv #clinicaltrials #externalcontrolarm #rwd #realworldoutcomes #patientjourney #realworldevidence #data #realworlddata #marketaccess #patientaccess #patientcentricity #outcomes #rct #hybriddesign #efficacy #evidence #patientoutcomes #medicalaffairs #clinicalpharmacology

56

Thanks to Kieran Gilmurray for describing the EU AI Act requiring businesses to train their employees about AI. It is exactly for this reason that I have published ‘AI Made Easy, A Beginner’s Guide to Artificial Intelligence’ to allow Everyone to have a really good basic understanding of large language models, algorithms and how AI is affecting their everyday lives including new AI systems coming into their work place. #AI #AIContent #ArtificialIntelligence #EUAIACT

8

#Ai realities and froth in #healthcare One of the The Labour Party ‘reform’ principles for the #NHS is analogue to digital. For digital one might hope it’s about the basic infrastructure and enabling rather than adding more work. It is definitely about the Ai elixir. Is the elixir real or just more snake oil? The best thing I read this week on anything Ai is the Substack transcript of the conversation between Eric Topal and a couple of seriously thoughtful experts, Arvind Narayan and Sayash Kapoor. Read or listen to it here 👇 https://lnkd.in/eeKJ5GJM The season for conferences is in full swing. Digital froth is everywhere. Seemingly good ideas and products are also slowly disappearing, recently described by Liam Cahill, a stalworth of digital realities in healthcare and things NHS. Is it a buyer problem or the solution assumptions or just timing? #Healthcare #digital

4

2 Comments

Harnessing AI to Combat Antimicrobial Resistance in Intensive Care The fight against antimicrobial resistance is more crucial than ever, with an alarming 1.2 million deaths occurring globally each year. This challenge has significant financial implications, costing the NHS at least £180 million annually. Artificial intelligence is now leading the charge in revolutionizing how we assess and manage these threats in intensive care settings. By providing same-day assessments of antimicrobial resistance, AI enables healthcare professionals to swiftly identify and respond to sepsis-causing bloodstream infections. Key metrics to consider: - AI can reduce the diagnostic time from five days to mere hours, enhancing patient outcomes significantly. - The interdisciplinary study involved data from 1,142 patients, paving the way for future research with datasets exceeding 20,000 individuals. - Targeted antibiotics can improve survival rates and reduce the development of further resistance, preserving current antimicrobial options. Join the conversation and share your thoughts on the role of AI in transforming microbiology and healthcare. Follow Spin Quantum Tech for more updates on innovations in the field! #Microbiology #ArtificialIntelligence #Laboratories #Future #Innovation #Health #QuantumComputing Spin Quantum Tech - SQT - Business acceleration through quantum computing. Dive into Quantum with Spin Quantum Tech: 📧: [email protected] 🏠: https://spinqtech.com 📱: +44 7537 188583

The open-source AI platform, Gemma, has launched its updated version, Gemma 2, in sizes of 9 billion (9B) and 27 billion (27B) parameters. Gemma 2 outperforms its predecessor with improved efficiency and safety. It offers a cost-effective alternative to larger models, operating on a single NVIDIA H100 Tensor Core GPU or TPU host. Built on a new architecture, Gemma 2 is compatible with AI frameworks like Hugging Face Transformers, JAX, PyTorch, and TensorFlow. It's optimized with NVIDIA TensorRT-LLM, and Google Cloud customers will soon be able to deploy and manage Gemma 2 on Vertex AI. The Gemma license allows developers to share and commercialize their AI innovations. The launch of Gemma 2 promises more challenging projects and improved performance in AI creations. The company plans to release a 2.6B parameter Gemma 2 model to address a wider range of AI tasks. Presently, Gemma 2 can be accessed in Google AI Studio, with model weights available for download from Kaggle and Hugging Face Models. Google also offers a free tier for Colab notebooks and Kaggle. Academic researchers can apply for the Gemma 2 Academic Research Program to receive Google Cloud credits for their research.

1

Insightful paper of the benefits of AI in stroke care https://lnkd.in/emj_UNcH

9