Will Medical AI Give Rise to the Next Alibaba in the Next Decade?

Regarding medical artificial intelligence, the most common refrain heard today is that it is “too early.” This is, in essence, a euphemism used by investors to justify not funding such projects. However, if we view this issue from a different perspective, failing to position oneself in the medical AI sector now could mean missing out on the opportunity to discover the next Alibaba over the coming decade.

Why Do Most People Consider Medical AI to Be in Its Early Stages? First, training artificial intelligence requires vast amounts of medical data, yet such data is highly sensitive due to patient privacy concerns. Second, medical issues are exceedingly complex, with information often lacking full transparency. Third, algorithms and data vary significantly across different diseases, resulting in a substantial workload.

These pitfalls have made it extremely difficult to develop mature medical artificial intelligence products. Moreover, no product on the market yet qualifies as AI in the standard sense, including IBM’s Watson for Oncology. Compounding this issue, the past five decades have seen repeated ups and downs in medical AI efforts with few breakthroughs, further fueling skepticism about its prospects.

Another aspect is that medical AI has not yet produced many compelling narratives. The most commonly cited application scenarios currently involve assisting doctors and hospitals in large tertiary institutions to improve the efficiency, quality, and safety of healthcare services; or providing decision support to primary care physicians in grassroots medical institutions, thereby enhancing the standard of care at the primary level.

Why are these stories unexciting? Because they are nearly identical to the narratives told when internet healthcare first emerged. Yet, to date, internet healthcare has brought only limited changes to traditional medical practice, although we remain confident about future transformations. After more than five years of exploration, people have increasingly realized that even technologies and business models mature in other sectors often fail to work in the healthcare field. The challenges in healthcare are too complex and profound; simply burning cash proves ineffective.

Coupled with the numerous existing pain points in healthcare, for which there are relatively clear improvement strategies and commercial prospects, this has prevented medical artificial intelligence—whose future remains uncertain and whose business models lack novelty—from becoming a favorite of investors. Not to mention the current investment winter; without proven monetization capabilities, startups may not even garner a glance from investors.

In general, given its unremarkable inherent capabilities and the intensely competitive landscape, it is perhaps only natural that medical artificial intelligence has fallen out of favor.

However, the view that medical artificial intelligence is still in its early stages is one-sided, or more specifically, static.

I, too, once believed that medical artificial intelligence was still immature and lacked clearly defined application prospects. However, after engaging with numerous healthcare startups dedicated to AI, I have revised my initial perspective. These companies are primarily concentrated inImaging Diagnosis, Assisted Diagnosisaspect. It is not to say that these companies have already produced flawless products, but rather that certain common trends emerging among these startups suggest that the potential for artificial intelligence to transform the future of healthcare is far greater than anticipated.

First and foremost, the common practice among these companies is actuallyDemonstrates the maturation pathway of medical artificial intelligence: Training algorithmic models with structured data. In the realm of algorithms, significant advancements have been made since the emergence of deep learning technologies, and a growing number of experts proficient in these techniques are coming to China. With regard to data, although challenges remain substantial, numerous practical breakthroughs have already been achieved. Coupled with China’s large population base, even localized data breakthroughs have the potential to drive significant progress in medical artificial intelligence.

Furthermore, medical artificial intelligence has alreadyBreakthrough Progress in the Application for Certain Disease Types. For instance, IBM’s Watson for Oncology has been widely applied in the auxiliary diagnosis of tumors. Furthermore, many domestic startups have made numerous attempts in medical image analysis and have successfully achieved progress in diagnosing pulmonary and cardiovascular diseases. In fact, most of these companies were founded only about two years ago. Such advancements are clearly remarkable.

These developments clearly demonstrate that the pace of artificial intelligence advancement may indeed be far faster than we anticipated. It took Watson less than five years to go from abysmal performance to winning the championship on Jeopardy!. Even with extensive training, humans lacking innate aptitude would likely struggle to achieve such a leap; however, the daunting aspect of machines lies in their unknown potential.

Beyond this algorithmic model, several key challenges that artificial intelligence still needs to address—including image recognition, speech recognition, and optical character recognition (OCR)—have become increasingly mature. Not long ago, Baidu showcased its “Baidu Brain.” Although Baidu has faced sustained criticism in recent times, the Baidu Brain, which merely possessed the intellectual capacity of a three-year-old just three years ago, is now widely applied across numerous fields such as mapping and autonomous driving, leveraging its technologies in image, speech, and text recognition. To borrow a theory from Kevin Kelly, the fundamental barriers to active learning in science and technology have been removed.

If we can build confidence in the technological advancements and breakthroughs of medical artificial intelligence, the next topic for further discussion is: where are the commercial application scenarios most likely to achieve breakthroughs first? On this point, we can begin by examining existingCommercialization Practices。

First is IBM’s Watson for Oncology. IBM is currently heavily promoting Watson for Oncology, which is primarily deployed in leading medical institutions, including Bumrungrad International Hospital in Thailand, Manipal Hospitals in India, and 21 hospitals in China such as the Sun Yat-sen University Cancer Center, the Fourth Affiliated Hospital of China Medical University, Zhejiang Provincial Hospital of Traditional Chinese Medicine, and Fudan University Shanghai Cancer Center.

One of IBM's path selections isRelated to the characteristics of tumors, as cancer patients tend to congregate at top-tier medical institutions; the other is related to Watson’s inherent characteristics. Since it primarily provides auxiliary diagnostic recommendations to physicians through literature retrieval, it is more adept at handling complex and rare conditions, which are evidently more concentrated in large hospitals.

The other is the path chosen by Baidu Medical Brain, which was just released, and a large number of medical artificial intelligence startups, namely focusing onApplied to Primary Healthcare Institutions. The fundamental concept of this approach is to train a mid-level medical AI assistant using data from numerous top-tier medical institutions, thereby providing primary care facilities with higher-quality second opinions.

This strategic choice was made after considering multiple factors. First, for AI products that primarily provide auxiliary diagnostic opinions, it is essential to integrate them seamlessly into physicians’ existing workflows; modifying the workflows of primary healthcare institutions is significantly easier. Second, the demand for auxiliary diagnostic support in large hospitals is far less urgent than that in primary healthcare institutions.

However, it is difficult to determine which direction offers better commercial prospects. Much like 16-slice CT and 256-slice CT scanners, products based on the same technical principles may target entirely different markets, yet both can generate substantial business.

Is the “next decade” timeframe just a wild guess? To some extent, there is an element of speculation. However, data from several areas can provide reference points for judging the length of this cycle.

The emergence of artificial intelligence represents a significant transformation in healthcare services, necessitating changes in established human behaviors and the acceptance of novel technologies. From this perspective, implementation is more readily achieved in newly established healthcare institutions.

For instance, new types of clinics are currently emerging on a large scale across China. Recent examples include DXY Clinic by DXY (Dingxiang Yuan) and Dr. Cui Yutao’s Clinic by Yuxueyuan. Beyond their innovative service models, these clinics are particularly impressive for their extensive adoption and application of new technologies to better serve patients. Given that the typical maturation period for general clinics is three to five years, it is reasonable to anticipate that new smart medical hardware devices will see widespread adoption within the next five years.

If we accept the hypothesis that new-type clinics are more willing to embrace innovations, then their large-scale growth and maturation will lay the foundation for the widespread adoption of medical AI products. However, this process will take at least five years. For larger hospitals, the maturation period requires 5–8 years, meaning the overall timeline could be even longer.

So, how long does it take to train an AI doctor? This is impossible to determine. Many startups have pointed out that, with current deep learning technologies, the process of training machines is broadly similar to that of training human physicians. Even excluding practical clinical experience, it typically takes eight years to train a qualified medical student. Although machines far surpass humans in data transmission and storage capabilities, the medical decision-making process is highly complex, making it difficult to estimate how long it will take for machines to replicate human decision-making and cognitive processes.

Another segment of the healthcare industry with the most mature value chain is new drug development. Typically, developing a new drug takes 10 to 15 years and involves several stages, including drug discovery, clinical trials, and regulatory approval for market launch. This process is highly analogous to the development cycle of artificial intelligence, which encompasses algorithm modeling, data training, and commercial deployment. In essence, the clinical trial process itself serves to demonstrate a drug’s efficacy through data.

These reference cases may not be entirely appropriate, but the development of medical institutions, the training of physicians, and the research and development of pharmaceuticals—all technical applications involving the human body—require a considerable “incubation period.” Therefore, it is likely that artificial intelligence technologies, which are also intended for direct application to the human body, will require a substantial amount of time to overcome these barriers. Of course, these figures could very well be nonsensical.

Another somewhat awkward issue facing medical AI is determining the actual size of its future market, which is particularly difficult to calculate. For instance, in well-known vertical domains such as diabetes, the population base and pharmaceutical expenditures are clearly defined, making the market potential straightforward to assess. The same applies to oncology, where market boundaries can be clearly delineated. But where do the boundaries of the AI market lie? While it is widely acknowledged that the prospects are significant, the exact scale of the market remains uncertain.

Here, we still refer to the conditions of several related markets.

First, an important application scenario for medical artificial intelligence products isComputer-Aided Diagnostic Analysis of Medical Imaging, collaboration with medical device manufacturers thus becomes a viable pathway to commercialization. In fact, among the CT equipment vendors we interviewed, many have recognized that the large-scale deployment of new imaging devices in China’s primary care market critically requires AI-assisted diagnostic and analytical tools, as the shortage of skilled personnel at the grassroots level significantly hinders market expansion. Therefore, the commercialization prospects can be considered relatively clear.

Analysis and forecasts regarding the scale of the medical device market are broadly as follows: China’s existing medical device market exceeds RMB 60 billion, with the medical imaging diagnostic equipment segment surpassing RMB 22 billion, making it the fourth-largest market globally. Forecasts indicate that by 2020, the size of China’s medical imaging diagnostics market will reach approximately RMB 40 billion.

The second market related to medical artificial intelligence isHealthcare Informatics, as many AI product applications are closely related to clinical quality and safety control, medical cost containment, health insurance audit, and personnel incentive management. These applications are all tightly linked to healthcare informatization. Currently, the healthcare informatization market is roughly characterized by the following: the hospital informatization industry scale is projected to reach RMB 24.82 billion in 2015, with a compound annual growth rate of 32.1%. Furthermore, research data indicate that China’s total spending on healthcare informatization is expected to reach approximately RMB 34 billion in 2017.

By contrast, the primary care market, where medical AI holds significant potential, may offer even greater promise.

According to the 2015 Health Statistical Bulletin, there were 7.7 billion patient visits throughout the year, with 6.2 billion occurring at medical institutions below the tertiary level. Secondary and lower-tier medical institutions, which require significant improvement in medical standards, will become the primary users of AI-assisted diagnostic systems in the future. These institutions already account for 80% of total patient visits. This is the source of the sensational headline of this article.

Primary healthcare institutions, which have become the focus of numerous startups, already handle nearly 4.4 billion patient visits. If China achieves its tiered diagnosis and treatment target—where 90% of patients receive care within their home counties—this figure is only poised to rise further.

However, what exactly does the vast primary care market signify?

China has improved its medical service delivery system, with efforts to strengthen primary care largely unfolding in three phases. In the first phase, the government used administrative measures to promote the downward deployment of healthcare personnel and upgrade infrastructure at the primary care level. This process is still ongoing. The most significant change in China’s primary healthcare institutions during this phase was a substantial improvement in hardware facilities, driven by considerable fiscal investments from governments at all levels. According to a set of healthcare reform data from 2013, capital construction investment for primary healthcare institutions amounted to RMB 130 billion, covering major equipment procurement, renovations, and maintenance.

The second phase involved the widespread application of internet technology within the healthcare service system, particularly in connecting medical institutions across different types and hierarchical levels. The significance of internet technology lies in reducing the cost of decentralizing medical resources through the use of telemedicine, and in establishing green channels for patient referrals via improved information connectivity. During this phase, regional healthcare information systems emerged in many areas. Moreover, the government demonstrated substantial financial commitment in this regard. In early 2015, the National Development and Reform Commission (NDRC) designated five provinces and autonomous regions as pilot sites for telemedicine, with total investment exceeding RMB 800 million.

Thus, from the perspective of the first two phases, the government has been highly willing to invest in advancing primary healthcare. The third phase may well involve the application of artificial intelligence (AI) at the primary care level. This is because the first two phases primarily achieved the decentralization of facilities and equipment, while the decentralization of personnel and technical expertise has yielded limited results. In other words, the issue of inadequate supply of high-quality medical resources at the primary care level remains unresolved. If AI can achieve breakthroughs in supplementing the supply of medical resources, government procurement alone could serve as a viable reference market.(for reference only, not government-led)。

Artificial intelligence is inherently compatible with the internet and smart hardware. Therefore, if medical AI can establish an application platform that is realistically scalable through large-scale standardization, the emergence of super-companies akin to Alibaba or Tencent in the healthcare sector may indeed be more than mere speculation.