Editor: Layan
What concrete and visible help will AGI have for us at the actual level? Presumably, we cannot escape the major innovation in the medical field.
Why do we need AGI?
Many people may not have thought about this question carefully and only saw the results. Some people believe that scientific and technological progress should be unconditionally promoted. As to the reasons for this, it may not have been thought through.
Perhaps in the end, it is only a conclusion that makes our lives more convenient. And what kind of convenience is provided in which aspects, it may not be so perfect.
In fact, this line of thinking cannot be wrong.
After all, some things are driven by motivation and then gradually realized. Some things are done first, and then see what can help us.
Today we will provide you with an entry point in a realistic context to see what AGI can help us.
The most comprehensive doctor
First, let's introduce one person, Greg Brockman.
Friends who are familiar with the AI circle should be no stranger to him, he is the co-founder + president of OpenAI.
From 2010 to 2015, he served as CTO at Stripe. Since then, he has been the president of OpenAI.
But today I'm going to talk about his wife, Anna Brockman.
In 2019, the two of them got married, and behind the happy marriage is his wife's physical condition that cannot be ignored.
In a recent tweet, Greg wrote, "After five years of experiencing multisystem pain in the body, my wife was recently diagnosed with a genetic disorder called hypermobility Ehlers-Danlos syndrome (hEDS). 」
Know that hEDS is an inherited connective tissue disorder that causes generalized joint overmobility, joint instability, and chronic pain.
At the same time, hEDS is accompanied by a variety of other symptoms and related conditions, affecting many different parts of the body.
Greg's wife is an actress for almost six years, and her current profession is a fitness instructor. One can imagine how much pain this disease will bring her.
Judging from the introduction of this disease, it is a comprehensive disease. It involves many systems in the body, such as orthopedics, cardiology, neurology, gastroenterology, dermatology, and so on.
Greg tweeted that the current healthcare system is built for each specialty. There are too many doctors to see for hEDS.
"In five years, we've seen more doctors and specialists than Anna had in her previous life. Most physicians focus on what they are familiar with and are not able to integrate the fragmented information. 」
Later, one of Anna's allergy doctors listened carefully to all of her symptoms and problems, piecing together the details of her physical condition.
Greg says that as human medicine advances, there seems to be a tendency to increase the depth of doctors at the expense of breadth. But for patients, what we need is enough breadth and enough depth, both of which are indispensable.
Ideally, in the future, we can make this comprehensive medical service pocket-like, like a team of experts composed of doctors from many departments to protect our health.
And that's where AGI comes in.
Greg concluded by tweeting that while there is still a long way to go on the technology side, and AGI has to learn how to deploy it together in a high-risk area like healthcare with the supervision of human experts, the outlook is getting clearer.
Through the collaboration of technology developers, healthcare providers, governments, and society, there is hope for a future where people can provide better healthcare for all family members.
Many netizens also expressed their opinions along the post.
Bacarella said that if medical AI can be as smart as the average doctor, and at the same time as patient, focused, and knowledgeable as GPT4, then it is estimated that there will be a lot of changes in the future.
Paul also believes that when AI develops to a certain stage, it will definitely be able to promote new medical diagnosis and treatment methods, so that the general public can access various medical technologies.
And this is probably the most important area for AI to devote itself to.
Google's attempt
You know, Greg's idea has been confirmed by Google before.
Researchers from Harvard University, Stanford University, Yale Medical School, University of Toronto and other top universities and medical institutions in the United States have jointly proposed a new paradigm of medical artificial intelligence, namely "general practice medical artificial intelligence", which can flexibly encode, integrate and interpret multimodal data in the medical field, such as text, imaging, genomics, etc., at scale.
Link to paper: https://www.nature.com/articles/s41586-023-05881-4
Google Research and Google DeepMind have also jointly published a paper to realize and verify the concept of artificial intelligence in general practice.
Paper link: https://arxiv.org/pdf/2307.14334.pdf
The researchers first curated a new multimodal biomedical benchmark dataset, MultiMedBench, containing more than 1 million samples and 14 tasks such as answering medical questions, interpreting breast and dermatology images, generating and summarizing radiology reports, and identifying genomic variants. Then, a new model, Med-PaLM Multimodal (Med-PaLM M), was proposed to verify the feasibility of a general biomedical artificial intelligence system.
It is a large, multimodal generative model that provides the flexibility to encode and interpret biomedical data, including clinical language, imaging, and genomics data, with only one set of model weights. Med-PaLM M is on par with state-of-the-art technology in all MultiMedBench tasks, and in some cases even surpasses dedicated SOTA models.
This paper also reports that the model can be generalized to new medical concepts and tasks, cross-task transfer learning, and emerging zero-shot medical reasoning ability under zero-shot learning.
The capabilities and limitations of Med-PaLM M were further explored, and the researchers compared model-generated and human-written chest X-ray reports for radiologists to evaluate, and in 246 cases, clinicians found Med-PaLM M reports to be better than those written by radiologists in 40.5% of the samples, indicating the potential clinical utility of Med-PaLM M.
To train and evaluate the ability of large models to perform a variety of clinically relevant tasks, researchers at Google collected MultiMedBench, a multitask, multimodal benchmark dataset for general practice.
The benchmark consists of 12 open-source datasets and 14 independent tasks, containing more than 1 million samples, covering multiple areas such as medical Q&A, radiology reports, pathology, dermatology, chest X-ray, mammography, and genomics.
Not long after, Google went on to launch Med-PaLM 2, the second generation of the product.
It is the successor to Med-PaLM, which is even more powerful than its predecessor, achieving an 86.5% accuracy rate on USMLE-style problems, an improvement of 19%.
Med-PaLM 2 is trained on massive medical text and code datasets, including medical journals, clinical trials, and textbooks. This allows it to understand and generate medical language with a high degree of accuracy.
Not only that, but Med-PaLM 2 is also capable of reasoning and inferring based on medical knowledge.
According to expert analysis, Med-PaLM 2 has the potential to revolutionize healthcare in several ways. For example:
Improve the accuracy of diagnosis: Med-PaLM 2 can help doctors comprehensively consider all of the patient's medical information, including symptoms, medical history, and test results, to determine the correct diagnosis for the patient.
Increased efficiency: Med-PaLM 2 can help physicians automate tasks such as summarizing medical records and finding relevant information from research papers. This frees up doctors to spend more time communicating with patients.
Improved communication: Med-PaLM 2 can help physicians communicate complex medical information to patients in an easy-to-understand manner. This can help patients make informed decisions about their treatment.
Cost Reduction: Med-PaLM 2 can reduce healthcare costs by automating tasks and increasing efficiency.
Currently, Google's Med-PaLM 2 is still in development, but it has the potential to have a significant impact on the healthcare industry.
However, when it comes to Google, I still have to mention the medical dialogue AI that just came out two days ago - AMIE, and it has directly passed the Turing test!?
Until now, Google is still actively testing it in an effort to make it more widely available in the future.
I wonder if products like Google, as well as other medical AI and even medical AGI that may appear in the future, can solve Greg's problem.
Resources:
https://twitter.com/gdb/status/1744446603962765669