AI medical treatment is coming to "knock on the door", is it the reincarnation of Jianghu Langzhong or Hua Tuo?

Text | Silicon-based laboratory, author: Bai Jiajia

The three key elements of good disease (quality), affordability (cost) and good health care (efficiency) have long been considered the "impossible triangle" in the medical field.

Under the wave of large models, can AI break this "medical impossible triangle"?

There are many optimists. Greg Corrado, a Google executive who has participated in the training of the large medical model Med-PaLM 2, said in an interview with the media that "AI can create 10 times the value in the medical field where it can bring benefits."

However, as soon as the perspective shifts to the patient or the healthcare worker, the attitude immediately becomes cautious.

In May last year, the WHO stressed in a statement that the integration of AI into healthcare services should be taken "very cautiously". Greg Corrado also said of his family, "I don't think it's at the point where I'd like to let my family use it."

Pull the horizon back home.

AI medical care has not yet been realized in the financial reports of enterprises, but the action of technology implementation is not slow at all. Since February last year, Internet giants have successively reached cooperation with hospitals and medical companies, launched large medical models, and landed model projects, and have been very busy.

It is worth mentioning that AI medical care is not a new species, and after the rise of the last round of AI with computer vision as the core, the story of AI diagnosis has been floating in the market. But over the years, the changes that ordinary people can feel are not strong.

The question then arises, will AI healthcare with generative AI at its core repeat the mistakes of the past? And how are the participants pushing it to a more pragmatic place? And how far are we from better AI healthcare?

01. The "own hand" and "wonderful hand" of AI medical treatment

AI medical care has been pinned on from the very beginning. In June last year, when most enterprises were still refining models, the Beijing Municipal Science and Technology Commission and the Zhongguancun Management Committee released the first batch of 10 large model application cases in the artificial intelligence industry to guide AI to the industry. Among them, healthcare is the only field that has been selected for two cases.

According to IDC, the global AI application market is expected to reach $127 billion by 2025, of which the healthcare industry will account for one-fifth of the total size.

So, what can AI medical care worth 100 billion yuan do?

At present, AI medical treatment mainly has four major directions: medical imaging, drug research and development, medical text processing, and academic research.

Among them, the two directions of drug research and development and academic research are currently limited by the small amount of data, so the progress is relatively slow. There are a lot of cases in medical text processing and medical imaging.

Among them, medical text processing is currently the most intensive direction for AI applications. With its strong information retrieval and text generation capabilities, AI can drastically reduce the repetitive work of doctors.

According to the example given by Huang Haifeng, general manager of strategic research and development of Baidu's big health business group, it used to take about an hour and a half or even two hours for doctors to write a discharge case, but with the assistance of AI, the time can be compressed to less than 20 minutes.

Medical imaging, on the other hand, is the direction with the strongest commercialization potential.

The potential mainly comes from two aspects, one is the combination of AI with CT and X-ray machines. Judging from past experience, hospitals have never been soft on the purchase of medical equipment and equipment.

The second is to facilitate the conversion from online consultation to actual transaction, for example, after users take photos to identify skin diseases, they further pay doctors for advice. According to the data disclosed by JD.com, the accuracy rate of AI-assisted diagnosis based on large models of JD Health Skin Hospital exceeds 95%, and the conversion rate of patients for special disease follow-up services has reached 20%.

Along the two major directions of text processing and medical imaging, most AI companies have formed a "large and comprehensive" AI medical layout, and are currently in full swing to integrate it into actual medical scenarios. In this competition, the big manufacturers have "their own hands" that must be done well, and there are also "wonderful hands" that are surprisingly successful.

The key to the so-called "own hand" is to work steadily and steadily to accumulate potential for the future.

All enterprises that want to make a difference in the field of AI healthcare cannot do without two points, data and tools. Data is the foundation of AI evolution, and the abundance of AI tools allows it to participate in the industry chain as much as possible.

Technology companies such as Baidu, JD.com, Tencent, and Huawei have established close cooperative relations with a number of medical institutions in the past year or so, and have promoted a variety of AI medical products.

For example, as early as February last year, before the launch of Wenxin Yiyan, Baidu announced the establishment of cooperation with a number of medical-related companies such as Moting Medical and Neusoft Medical. At present, five major products have been formed for various groups in medical scenarios, namely AI health assistant for ordinary users, online medical Copilot for doctors and patients, "AI smart clinic" for hospitals, CDSS (clinical decision support system) for large-scale private products, and Lingyi open platform products for enterprises.

Among them, CDSS has been implemented in more than 4,000 primary medical institutions because it can effectively improve the diagnosis and treatment ability of grassroots doctors and reduce misdiagnosis and missed diagnosis.

In addition to the "own hand", it is worth paying more attention to the "wonderful hand" of the big manufacturers. These "magic hands" come from the company's own advantages, and strive to stand out in the homogeneous competition.

For example, JD.com has taken advantage of its logistics system to move online consultations to offline.

On the basis of AI consultation, JD.com recently announced a new upgrade to the patient service of JD Health Internet Hospital. If there is a need for further examination during online diagnosis, you can place an order through JD.com's home quick test, and a nurse will come to collect blood, sample and send it for testing, and the results will be available in an average of 3 hours.

Huawei, on the other hand, has taken advantage of its two-legged approach to software and hardware to bring its AI healthcare business to the upstream of the industrial chain.

The Silicon-based Lab found that in the cooperation between Huawei and Yidu Technology, Runda Medical, Jinyu Medical, and Wanda Information, in addition to launching a large model of the medical examination industry, it also includes an all-in-one machine for medical models based on Ascend AI's basic software and hardware, providing customers with out-of-the-box AI full-stack solutions.

On the whole, since the development of AI healthcare, the industry's attention has not been limited to the strength of technical capabilities in a certain link, but has paid more attention to whether a specific landing plan has been formed and whether users can get a better experience from it.

However, compared with the burning oil in the industry, investors still maintain a certain cold thinking.

"I don't think that the medical model is currently an outlet, the large model is just a capability, and the progress of technology is still very long to really change an industry, especially in medical care." Sun Motao, vice president of Qiming Venture Capital, previously revealed in an interview that after the disenchantment of AI, the investment circle is still more practical about the medical model.

02. "Multimodality" may be an opportunity for overtaking

As He Mingke, president of Baidu's big health business group, said, medical care is a rare "three-way separation" industry: the decision-makers are doctors and hospitals, the payers are insurance and medical insurance, and the users are patients.

The complex chain of "decision-payment-use" and the seriousness of the medical industry have made the implementation of the medical model more rigorously scrutinized than other industries, so it has gradually shown a trend from light to heavy in the process.

From light to heavy, that is, it is first applied in fields that are easy to land, have strong user perception, and are not easy to cause medical accidents, and then advance to more subdivided scenarios.

It is worth mentioning that on the one hand, it is based on the company's own hedging logic, and on the other hand, it is also a hard rule. In September last year, the Beijing Municipal Health Commission led the formulation of the "Beijing Municipal Implementation Measures for the Supervision of Internet Diagnosis and Treatment (Trial)", which clearly prohibits AI from conducting diagnosis and treatment without the supervision of a doctor, or automatically generating prescriptions.

The light-to-heavy trend is mixed news for practitioners. On the one hand, the technical threshold for entering the game is not high, but in view of the particularity of the medical field, it also means that it will be difficult for rising stars to have the opportunity to surpass those large manufacturers with money and "connections".

Specifically, medical data is relatively sensitive, and changing partners in the future may have a greater impact on hospital operations, so when initially selecting partners, they will prefer large manufacturers with guaranteed reputation and risk control capabilities.

Therefore, considering the dislocation competition with large factories, small factories often use multimodality to achieve corner overtaking.

Compared with single-modality, multimodality involves a wider range of technologies, and multimodality is the only way for AI medical care. Just as doctors need to accumulate information through multiple examinations before giving a diagnosis, AI also needs to integrate multiple forms of information to make accurate judgments.

For example, the premise of a multimodal AI is that it can find anomalies in the information. Compared with generative AI, whether the final conclusion is correct or not depends to a greater extent on the detection ability of vision AI.

In other words, if a company can demonstrate multimodal capabilities that can "win in the future", hospitals are also likely to cooperate with it based on long-term stability considerations.

The market has also proven the appeal of multimodal AI. During the 2024 World Artificial Intelligence Conference, SenseTime, together with Ruijin Hospital, West China Hospital and other hospitals, took the lead in launching an innovative demonstration project of medical multimodal AI-empowered smart hospitals in the industry.

According to public information, all parties will use SenseTime's large language model as the "central brain" of the smart hospital to intelligently dispatch a special model covering various data modalities such as medical texts, radiological images, and pathological images to assist doctors in completing complex diagnostic reasoning across departments and modalities.

03. How far away is AI medical treatment?

If you want to give the steak name that the tech giants have boasted about, AI medical can probably rank in the top three. When the wave of deep learning rose in 2012, Hinton, one of the three giants of deep learning, once said that AI could replace radiologists within 5 years.

But more than 10 years later, the AI has not been as he hoped, and Hinton is clearly overly optimistic.

So, will this wave of AI medical care be as "thunderous and rainy" as before?

At present, there are mixed joys and sorrows.

The "good news" is that some of the technical difficulties of the past are indeed being slowly overcome by generative AI.

The lack of high-quality datasets has always been a big problem for AI companies, especially in medical scenarios.

For example, there is enough data for common diseases, but it is difficult for rare diseases and genetic diseases to meet the information requirements, even if the traditional AI training paradigm of "big data + fine labeling" is adopted, because each doctor writes a report and has different habits of using medical devices, it is also an almost impossible task.

However, generative AI can alleviate this problem, for example, a study by Shanghai Jiao Tong University used the understanding of medical images and concepts contained in medical language models to guide the migration of visual models trained by natural natural images to case images, so as to complete the task of low-sample classification of pathological images.

This type of technology is already being used in the industry. PathOrchestra, a large pathology model recently released by SenseTime, learns to analyze pathological images of various organs through self-supervised learning of massive data without the need for a large number of fine annotations.

The "worry" lies in the fact that these companies still have difficulties to overcome in the process of commercialization, the most important of which is the problem of capital.

From the standpoint of enterprises, although the current competition in the field of AI medical has moved from technology to landing, the driving force comes more from the demand to build model projects in various places, rather than profit-driven expansion.

In the absence of mature solution integrators, this trend puts forward high requirements for the lead enterprise, which not only needs to have very strong capital advance and management capabilities to integrate the upstream and downstream, but may even need to continue to provide a dedicated service team for relevant parties after completion.

At the same time, the seriousness and professionalism of AI medical care determine that companies are unlikely to dilute costs through rapid replication in a short period of time, and with the already high R&D investment of such technology companies, companies that are not optimistic financially or large enough are likely to fall on the road.

Although the financial pressure is huge, these companies have no other choice, and if they want to truly meet the needs of doctors and patients, AI must enter the real medical scene for iteration.

As Liu Hongbin, executive director of the AI Center of the Hong Kong Innovation Institute of the Chinese Academy of Sciences, said in an interview: "Many large models do not involve hospitals in the research and development process, and lack of clinical factors, resulting in some seemingly cool technologies that actually do not meet the needs of doctors." ”

Switching the perspective to hospitals, leaving aside those model projects, in fact, "money" is also a big problem.

At present, most hospitals mainly use CPUs for computing resources, and few GPUs are required to deploy AI. Therefore, if you want to implement AI healthcare, you need to be equipped with GPUs at the same time, and ensure sufficient storage and high-speed network connection.

At a time when there is a global shortage of chips, this is a cost that many hospitals can hardly afford. According to the estimation of one H20 for one department, on average, each hospital needs to invest millions of yuan in chip configuration.

From this point of view, if AI healthcare companies want to achieve the large-scale implementation of new technologies, they must make a killer application, so that hospitals can willingly upgrade the deployment environment in order to purchase AI.

On the whole, although the medical model has begun to land in some scenarios, it has not yet reached a turning point that can really change people's lives. After all, it's hard to get people to trust AI to see a doctor.