After two amazing years @GoogleDeepMind, I’m now joining @OpenAI to accelerate biomedical intelligence with @thekaransinghal and team! 🧬🩻 Incredibly lucky to have worked alongside @alan_karthi @vivnat @taotu831 @RyutaroTanno @valentinlievin and many others to develop AMIE and Med-Gemini. Excited to continue my mission on building safe AI that benefits humanity 🚀

Introducing Med-Gemini, a family of models that extends the best of Gemini into medicine! ✨⚕️ Highlights of what you can do with Med-Gemini: > Answer medical questions with up-to-date knowledge using agentic web search 🔎❤️‍🩹 > Converse about your medical images, videos, and long multi-visit health records 📷📹📃 > Do a literature search by uploading tens of biomedical papers and asking questions 📚 > And so much more! 🏗️ Development of Med-Gemini included: > Advancing clinical reasoning with self-training and search > Improving multimodal understanding with fine-tuning > Leveraging long-context capabilities with chain-of-reasoning Paper: arxiv.org/abs/2404.18416 Below ⬇️, I talk more about self-training with web search to improve Gemini’s clinical reasoning.

Gemini powers our multimodal health research! 💙 In our new paper on multimodal AMIE, we're pushing conversational diagnostic AI beyond text to handle images such as skin photos, ECGs, and clinical docs, which provide crucial context in healthcare. Blog: goo.gle/42D0QcB Paper: gstatic.com/amie/multimodal_… How do we make an AI reason like a clinician during a dynamic, multimodal conversation? One of our key contributions is multimodal state-aware reasoning, built on @GoogleDeepMind Gemini 2.0 Flash. Instead of just reacting turn-by-turn, AMIE maintains an internal "understanding" of the consultation: ✅ What is known about the patient? ✅ What are the likely diagnoses? ✅ What information (text or visual) is missing? This internal state allows AMIE to: 👉 Intelligently guide the conversation through phases like history-taking & diagnosis. 👉 Strategically ask for relevant images (like skin photos or screenshots of ECGs/docs) when its internal state shows uncertainty. 👉 Accurately interpret multimodal data and weave the findings back into the ongoing dialogue and diagnostic process. Essentially, it mimics the adaptive reasoning clinicians use, leading to a more structured and effective consultation. We evaluated multimodal AMIE against primary care physicians (PCPs) in a demanding, blinded OSCE study using 105 diverse multimodal scenarios. The results demonstrate clear progress: AMIE achieved similar or superior performance when compared to PCPs across a wide range of metrics, including diagnostic accuracy, empathy, and critically, the handling and reasoning about multimodal data. While the OSCE results are very promising, it's important to remember this was a test environment with patient actors! Real-world care is more complex. Making sure it's safe, reliable, and actually helpful in the real world needs more work, starting with our upcoming study with Harvard BIDMC. The work would not have been possible without an amazing team @GoogleAI, @GoogleDeepMind: @RyutaroTanno, @alan_karthi, @vivnat, @AdamRodmanMD, @timstro, @taotu831, @hardyshakerman, @JanFreyberg, @_cjpark, @yasharmaa, @apalepu13, @arkitus, @weballergy, @valentinlievin, @ckbjimmy, @davidstutz92, @dgtbarrett, @yongcheng16 @SaraM66905, @dr2w, @ymatias

My final PhD chapter on improving seizure detection with @HazyResearch and @rubinqilab was just published @npjDigitalMed. TL;DR We found that scaling two dimensions of model supervision: (1) coverage of training data and (2) granularity of class labels– has a large impact on model performance AND subgroup robustness for seizure detection. The best part? We do it using freely available notes produced in routine clinical workflows! Clinical importance of building reliable seizure detection tools. Detecting seizures, their types, duration, etc. is a critical healthcare task in diagnosing and managing epilepsy. The best way to do that is EEG analysis (reading brain recordings). But EEG analysis is a BIG pain! These recordings can be very long (hours-days per patient) and demand a scarce resource: deep neurologic-epileptologic expertise. So we have a strong need to develop reliable tools to help clinicians analyze EEG more efficiently. Why aren’t existing models widely used? A big reason is trust. ML models often fool us on aggregate metrics, where they show expert-level performance on average and then, whoops, they turn out to rely on non-causal features and do a lot worse for certain subgroups. We can’t have that for high-stakes healthcare settings. Another major reason is high false-alarms on abnormalities that may look like seizures, leading to alarm fatigue. Workflow notes: a hidden goldmine for supervision. Standard seizure detection models rely on manual labels from experts, but this approach is too expensive to scale. Luckily, routine clinical EEG monitoring leaves a trail of helpful annotations from techs, fellows, & docs. These workflow notes provide an opportunity to freely scale supervision for seizure detection models. Scaling coverage is not enough. Using workflow notes, we scaled our training data to include ~70k hours of EEG from ~12k patients. While this gave us impressive overall performance, we found significant performance gaps among certain patient age groups and seizure subtypes. We also found many false positives on non-epileptic abnormalities. Scaling granularity of class labels is also needed. Since workflow notes also include events beyond seizures (e.g., spikes, patient movement), we trained a multilabel model to predict 26 classes (including seizure). The intuition here is that increasing class granularity teaches the model to differentiate between seizures and other non-seizure abnormalities, lowering false positives. We found that our multilabel model improved overall performance, and importantly, had no significant performance gaps among subgroups. Concluding thoughts on supervision. It’s amazing how supervision has such a large impact on model reliability. Since supervision in healthcare is scarce, we should always keep an eye out for how we can leverage existing routine workflows to supervise our models – I am sure many more exist that we aren’t taking advantage of yet! — This work was done alongside amazing collaborators @SiyiTang_, Mohamed Taha, and @ChrisLeeMesser It was also inspired by earlier explorations with @jdunnmon and @ajratner And was supported by @StanfordBrain and @StanfordHAI Check out the paper for details, including how we used SSMs for modeling, and improved on clinical utility metrics. nature.com/articles/s41746-0…

LLMs are becoming more capable in health, and we need to think of safe ways to make use of them. In our recent AMIE work, we envision using LLMs to directly scale the physician. We imagine a setting where a physician deploys tens-hundreds of LLMs to gather patient information for them. These LLMs (guardrailed not to give medical advice) have conversations with patients async, and come back with structured summaries so that the physician can review, edit, and submit the appropriate medical advice. We ran a simulated study with patient actors to safely test this idea, comparing AMIE against nurses and physicians. Check out the blog and paper to learn more! research.google/blog/enablin…

Fun to hear Google SVP James Manyika at IO mention our work on Med-Gemini as one of the biggest opportunities and applications for AI. Exciting things to come 🚀

Our results using AMIE are pretty exciting! But how did we specialize AMIE to have high quality diagnostic conversations? Let’s talk about our core training recipe: the *simulated dialogue learning environment* 🚂 Preprint: arxiv.org/abs/2401.05654

Our approach to evaluating health AI models continues to evolve! (Phase 1) Medical Benchmarks ➡️ (Phase 2) Patient Actor Consultations ➡️ (Phase 3 – coming soon!) Real-World Deployment > (Phase 1) Medical Benchmarks: We first need to make sure our models have extensive medical knowledge and clinical reasoning ability. To do this, with Med-PaLM and Med-Gemini, we focused on evaluating our health AI models on a broad range of medical exams and established benchmarks [1]. > (Phase 2) Patient Actor Consultations: We then need to take a step closer to real-world deployment. With AMIE, we ran studies with patient actors to understand how our health AI models perform in more realistic, conversational settings [2]. > (Phase 3) Real-World Deployment: Coming next, in partnership with Harvard @BIDMChealth, we're launching a prospective study to evaluate AMIE in a real-world clinical setting, exploring its impact on pre-visit information gathering and the perceptions of both clinicians and patients, under strict medical supervision [3]. Excited to report back what we learn! Check out our @GoogleAI blogs below for more info. [1] research.google/blog/advanci… [2] research.google/blog/amie-a-… [3] research.google/blog/advanci…

Checkout our work in @Nature_NPJ Digital Medicine with @jdunnmon @rubinqilab @HazyResearch and @cleemesser on sustainable training for automated seizure detection models using weak labels! 1/3 rdcu.be/b3Fo8

Having a "ground-truth" answer in medicine is not always feasible! As part of Med-Gemini, we did a deep-dive into a popular medical benchmark: the MedQA (USMLE) test set -- and found that 7.4% of question/answer pairs are ambiguous. Check out @davidstutz92's thread for the details and release. What do we do about it? We're not totally sure it's solvable. But, we take some steps to mitigate: (1) evaluating on a plethora of benchmarks (2) qualitative analysis with clinician feedback (3) designing new evaluation paradigms like we did with AMIE (double blind OSCE study with patient actors) (4) more to announce soon!

How can you reduce reliance on spurious features? At MLHC, we show you can use ~spatial specificity~ (1/7) > paper: web.stanford.edu/~ksaab/medi… > video: piped.video/watch?v=9uDsWp-C…

Early detection of neurodegenerative disease has been a major challenge, often requiring invasive tests or waiting for clear symptoms. The stellar team @PrimaMente are tackling this challenge head on by building epigenetic foundation models that analyze DNA fragments in the blood (cfDNA) to find early signs of Alzheimer’s and Parkinson’s. 🧠🧬 Fortunate to have advised on this project in a personal capacity.

Behind the scenes look into the making of AMIE, Med-Gemini, and the exciting future explorations in medical AI at @GoogleDeepMind. Thank you @labenz for hosting @vivnat and I on The Cognitive Revolution podcast and for the stimulating discussion!

Medical AI has a lot to benefit from advances in long-context reasoning ⚕️📚 In our most recent work with AMIE -- our research project on conversational diagnosis & management -- we leveraged Gemini's long-context reasoning to (1) generate personalized management plans grounded in trusted clinical guidelines, and (2) perform longitudinal reasoning on disease progression across multiple visits. We ran a study to compare AMIE's new skills against primary care physicians in 100 multi-visit scenarios with patient actors. Check out Valentin's post, the blog, and paper to learn more! research.google/blog/from-di…

Our work on AMIE is published in Nature! Exciting follow-up studies coming soon 🏥

AMIE helps general cardiologists assess patients with rare inherited heart conditions 🫀🧬 Getting a diagnosis for a rare heart condition is challenging, especially when it involves looking at results from a bunch of different tests like electrocardiograms, echocardiograms, cardiac MRIs, genetic tests, and cardiopulmonary stress tests. We wanted to see if AMIE – our research LLM specialized for diagnostic conversation – could help general cardiologists in complex sub-specialty cases for cardio-genetics. To enhance AMIE’s responses in this sub-specialist domain, we used few-shot prompting, web search tool-use, and self-critique. Then, we had sub-specialist cardiologists use a new 10-point checklist to compare AMIE's assessments with those of general cardiologists. We found that AMIE did better than the general cardiologists in 5 out of 10 areas, and excitingly, improved the responses of general cardiologists when they had access to AMIE’s responses! These results highlight the potential for LLMs like AMIE to democratize subspecialist-level expertise and improve healthcare delivery. Future research will focus on further validation and integration of multimodal data. Paper: arxiv.org/abs/2410.03741

As we continue to see step changes in base model capabilities, our ability to advance human knowledge and unlock scientific progress accelerates 🔭🧪🩻

Really enjoyed talking about AMIE on the Stanford MLSys Seminar with @apalepu13! Thank you for hosting us @mzhangio @bfspector @Avanika15 piped.video/watch?v=KKasKOQA…

Our new virtual collaborator, AI co-scientist, loves exploring questions that scientists are curious about. When scientists share their research goal, it will gladly synthesize extensive literature, engage in internal debates, organize tournaments, and integrate expert feedback to iteratively refine research plans. For a great overview, and information about the trusted tester program, check out Vivek’s thread!

Congratulations @_albertgu @krandiash and team! I was lucky to learn from you during my PhD and apply the early technology to health applications. Excited for what you build next 🚢🫡

Really enjoyed collaborating with Anita Rau, @yeung_levy, and the Stanford crew on this! If we want AI to help in the OR, we need to know what today's VLMs can (and can't) do. Great to see Med-Gemini benchmarked alongside others here in this much-needed, rigorous eval for the surgical domain. Important step forward!

Clinical reasoning is an iterative process where physicians combine their knowledge with patient information to form a case representation, guiding further data collection until a diagnosis is confirmed. Importantly, as medical knowledge rapidly evolves, clinicians also integrate up-to-date information from authoritative sources. Lack of reasoning traces for supervision is a challenge when training LLMs to improve their clinical reasoning. A popular dataset for medically fine-tuning LLMs is the MedQA (USMLE) train set – but this dataset only contains answer choices to questions. Also, how do we train an LLM to effectively integrate information from sources on the web? Self-training with web search, where we use a handful of expert reasoning traces (with and without web search integration) to kickstart a self-training loop, alleviates the need to have large scale reasoning supervision. For each training question, Med-Gemini generates reasoning traces using the expert demonstrations as in-context examples. We then use the reasoning traces that got to the correct answer to self-train Med-Gemini. Uncertainty-guided search at inference is how we decide when to invoke search at inference. We ask Med-Gemini to answer the same question multiple times, and estimate uncertainty by calculating the entropy across the predicted answers – high uncertainty invokes an iteration of web search. We also found that it was important to generate search queries by contrasting Med-Gemini’s conflicting answers. In conclusion, we found that self-training with web search greatly improved Med-Gemini’s clinical reasoning ability, and led to a new SoTA on MedQA (USMLE), and generalized to other challenging benchmarks. Check out the paper for more details. And check out @taotu831 @RyutaroTanno @ckbjimmy for multimodal and long-context capabilities! nitter.app/taotu831/status/…

Excited to continue making progress towards the grand vision of democratizing access to medical expertise.

Check out our new time series modeling approach w/ SpaceTime! 🌌 We build on the exciting direction of embedding signal processing principles within deep learning models to achieve: ▶️ expressive modeling ▶️ efficient training + inference ▶️ long context + horizons

Check out @_albertgu's exciting work on sequence modeling! This is particularly exciting to me since many medical time series applications, such as seizure classification from EEG, may benefit when we better model long-range dependencies.

We aim to improve automatic discovery of underperforming slices by leveraging cross-modal embeddings! We're inspired by the growing work revealing poor performance on hidden stratifications in healthcare (e.g., checkout @DrLaurenOR work!)

I've really enjoyed using Meerkat. Since i'm dealing with multiple data modalities (images, reports, gaze data, etc.) it provides an easy way to slice and dice my data across modalities during error analysis. Highly recommend!

Check out our findings on how AMIE did in the specialty of breast oncology!

Congrats, great work! I especially like the consideration of healthcare cost. Really cool to see Dx accuracy vs cost curves. Also nice to see the framework is agnostic to the base LLM.

An exciting and very important problem especially for medical applications!

We find that CNNs trained using massive numbers of these weak annotations can achieve performance levels competitive with -- and in some cases, even surpassing -- deployed clinical software. 3/3

It was so fun to learn from and work alongside amazing mentors and teammates @vivnat @alan_karthi @taotu831 @RyutaroTanno @Mysiak @dgtbarrett @HardyShakerman @3llery @weballergy Christopher Semturs and so many more folks across @GoogleAI & @GoogleDeepMind. It was an incredible honor to be supported by our inspiring leaders @JeffDean @demishassabis @OriolVinyalsML @koraykv @ymatias @greg_corrado @joelle_barral. So excited to continue on the journey of investigating the art of the possible in health AI!

As many others pointed out in their great work (@jdunnmon @DrLaurenOR), we believe this is due to task underspecification. The binary classification task is not specific enough, so the model gets away with taking shortcuts. (3/7)

Inspired by works on synthetic data, we developed a multi-agent data generator for diagnostic conversations. Key concept: instruct one agent to act like a doctor, instruct another agent to act like a patient, let the agents chat! But how does this solve the two challenges?

(*Patient vignette generator*) To improve scalability, we scraped the internet for thousands of conditions, and leveraged web searches to retrieve symptoms, medical history, etc. associated with the conditions. We use these vignettes in the instruction to the patient agent.

(*Outer Self-play*) Combining the vignette generator with the patient, doctor, and critic agents, we generate diverse & high-quality diagnostic conversations, which we then use to fine-tune AMIE. Repeating this process improves AMIE’s diagnostic conversational abilities.

As opposed to expensive gold-standard labels provided by clinicians, weak labels are commonly produced in large volumes within clinical workflows by a mixed group of technicians, fellows, etc., which are often used by a clinician as a starting point for their own analysis. 2/3

So many awesome use cases for Meerkat! My favorite one: expanding and speeding up error analysis iterations

Contributing to building AMIE has been a thrilling experience. The best part of course was getting to work with amazing collaborators across @GoogleAI @GoogleDeepMind. So excited to continue pushing the boundaries of research in health AI here!

Thank you @BasantLashin1!

I want a refund on my California sunshine tax 🌧 #RefundSunshineTax

This work was only possible because of awesome collaborators Sarah Hooper, @MayeeChen, @mzhangio, @rubinqilab, @HazyResearch, as well as @StanfordHAI for supporting us with computing resources. (6/7)

Congrats Eric! So excited for your next adventure 🦾

Inspiring to see the school I started at @LebAmUniv making leaps forward in research output! Proud of the bright faculty/students that accelerate progress despite economically challenging times 🙏 news.lau.edu.lb/2022/lau-str…

That’s amazing @MichaelWornow Why do we bother with those salads when you visit Google 😅

Thanks. The study with actors is a safer more controlled setting before testing with real patients. We’re currently running a study with real patients but it takes longer.

Although we can transcribe real-world dialogues from visits, we face 2 challenges: 🏔️ Scalability: it’s challenging to capture numerous medical scenarios ✅ Quality: real-world conversations are noisy! They contain ambiguous language, interruptions, implicit references, etc.

(*Inner Self-play*) To improve quality, we introduced a third agent: the critic! We instruct the critic to give feedback to the doctor agent for self-improvement. The doctor integrates the critic's feedback in future rounds of dialogue.

Amazing work! Other than the tokenization piece, it feels like we now have a deeper understanding on how to mix SSMs with transformers compared to earlier hybrid approaches. I wonder how H-Net compares to pure forms of SSMs for signals data, like audio/EEG. Also feels like H-Net is better suited for modeling multimodality since dealing with modality-specific encoders/tokenizers adds complexity.

Med-Gemini models are research models to help us understand what's possible (and won't be open source), but stay tuned for announcements from Google Cloud's MedLM product offerings! cloud.google.com/blog/topics…

Much needed, especially in the medical domain! Congrats!

This is why it takes longer. The study with actors is a stepping stone.

It was my first time @mlforhc, and it was a blast! Met lots of folks who are also passionate about collaborating with clinicians to successfully bring ML to healthcare! (7/7)

Academia 🤝 Industry An important perspective on the AI race by my phd advisor

Congrats boss!!

We discuss the cost-robustness trade-off of spatially specificity and explore semi- and self-supervised learning techniques to reduce the cost of increasing spatial specificity -- we can get large robustness gains with just a few hundred segmentations! (5/7)

Thank you Ali :) Med-Gemini models are research models to help us understand what's possible (and won't be open source), but stay tuned for announcements from Google Cloud's MedLM product offerings! cloud.google.com/blog/topics…

Spurious correlations occur in many real-world datasets, and could be a major problem in healthcare ML. E.g., a model may show expert-level performance on classifying pneumothorax... but perform at random on CXRs without treatment tubes. (2/7)

We find a strong positive relationship between robustness to spurious features and the degree of spatial specificity of the task (e.g., segmentation is more spatially specific than binary labels, which improves model robustness). (4/7)

Congratulations!

Beautiful work!

Of course, we’re always happy to chat. We have our emails in the Med-Gemini paper arxiv.org/abs/2404.18416

My #1 job search advice: listen to @labenz podcast 🤣

Congrats!