Artificial intelligence is making waves across every industry, and in 2024, its integration into healthcare is rapidly becoming a defining trend.
Wang Xiaochuan, founder of Baichuan Intelligence, has confidently stated that healthcare is the sector his company understands best and finds most exciting. Baidu Health, in partnership with Wuhan Union Hospital, has launched a project aimed at enhancing outpatient services. Meanwhile, Huawei Cloud’s Pangu model is setting its sights on traditional Chinese medicine, and Tencent Health sees AI as the solution to the three major challenges in hospital management: patient services, doctor diagnosis, and smart administration.
At the recent “AI For Health” forum during the CPEO, several leading tech giants showcased their visions. With the potential for AI models to transform and upgrade the health industry, everyone is vying for the lead in the next decade of innovation.
Since 2015, during the previous AI boom, advances in deep learning, natural language processing (NLP), and computer vision reduced AI production costs and lowered barriers to entry. This attracted nearly RMB 1 trillion (USD 140.3 billion) in investments, sparking rapid growth in areas like AI-driven drug discovery, diagnostics, and medical imaging. The influx of capital also led to the rise of major players such as Alibaba, Tencent, and iFlytek, alongside newer medical AI companies like Yidu Tech and Airdoc.
However, this investment frenzy has brought significant issues to light. Few companies have successfully integrated AI technology into real-world healthcare settings. The general public remains largely unaware of these advancements, and the anticipated value of AI in healthcare has yet to materialize in financial reports.
As the industry enters its next phase, the focus is on applying AI effectively in the right scenarios. Could the new generation of generative AI models offer a fresh approach?
Technically speaking, early AI models were primarily task-oriented, requiring the definition of independent tasks, data labeling, and the selection of appropriate algorithms from a limited set, followed by extensive parameter tuning. Only the largest internet companies could afford the time and resources needed to develop applications during this period.
Around 2013, technological advances led to a unification of algorithms, significantly lowering AI generation and application costs. This ushered in the first wave of AI democratization. In healthcare, this period saw the rise of AI systems capable of performing specific tasks, such as assisting with diagnostics, analyzing medical images, and aiding in drug development.
Today’s AI represents a significant leap forward, achieving what experts describe as the complete unification of data, models, and tasks. The key difference lies in the advanced capabilities of modern large models to understand language, reason, and generalize, enabling them to handle natural language more fluidly and generate high-quality text.
“The intrinsic value of large models lies in their ability to generalize,” said Zhang Fan, COO of Zhipu AI, in an interview with 36Kr. “For example, in medical case identification, traditional independent models might require thousands of annotations, and any change in the case record format could invalidate those annotations. Large models, however, think more like humans—they might only need to learn from a dozen cases to transfer knowledge. As long as the information exists in their past training data, they can provide reasonably accurate answers.”
This approach leads to a significant difference in product capabilities. Traditional NLP systems might produce accurate text analysis, but they often struggle with natural language interactions in open scenarios. In contrast, large models can ask relatively accurate questions, find relevant knowledge, and generate personalized recommendations. Additionally, since large models can learn from publicly available knowledge, they require less human maintenance, reducing development costs.
In theory, these advances blur the lines between human and AI applications, ushering in a new era of human-AI collaboration. When it comes to application development, discovering hidden data links or reducing repetitive manual labor naturally becomes the first choice for large models in healthcare.
During the CPEO, various companies presented enterprise-level applications in traditional scenarios, such as smart consultation assistants and conversational data analysis. For example, Cheetah Mobile and Zhipu AI are introducing large models into the academic marketing processes for pharmaceutical representatives, facilitating real-world dialogue training between reps and doctors and generating visit summaries.
Yet, these applications often seem like merely reapplying new technology to old scenarios. From the end-user perspective, these are functional upgrades driven by technological advancements. For instance, Baidu’s new system may reduce the time doctors spend writing case notes—from an hour and a half using traditional AI to less than half an hour with large models. But can this really address deeper pain points? Should we expect large models to explore even newer applications?
Zhang candidly admitted that current large model products are “useful, but not revolutionary.” He notes that the emergence of truly transformative applications will take time. “The problems today’s AI addresses might seem similar, but the underlying technical logic has completely changed. This enables us to apply AI capabilities faster while reducing costs. Moreover, as model versions iterate rapidly and modalities increase, the range of applications will continue to expand,” Zhang said.
As of October 2023, nearly 50 medical large models had been publicly released in China, accounting for about 20% of all large model products. Functionally, there are already signs of overlap, with most models focusing on areas such as doctor assistants, patient consultations, and drug development—leading to a lack of differentiation in application scenarios.
This also means that, after initial explorations, large model companies now face a critical task: making money.
An industry insider told 36Kr that last year’s market was focused on model capabilities, but this year, it’s all about business. “Almost all customers now care more about how to translate model capabilities into business value. Beyond strong model capabilities, a complete service system, a deep understanding of healthcare industry know-how, and the ability to implement these solutions systematically are all becoming key evaluation criteria for buyers.”
At the CPEO, some large model companies showcased meaningful collaborations with pharmaceutical firms. For example, Huawei Cloud is exploring the use of its Pangu model for traditional Chinese medicine R&D. According to Wang Wenjia, general manager of Tasly’s international innovation center for gene-based drugs, the digitization of traditional Chinese medicine is a direction with significant demand but few solutions. The two parties have collaborated to develop an R&D model that, by learning from extensive repositories of knowledge, formulas, and treatment records, and fine-tuning based on the R&D scenario, can help achieve functions such as smart consultation and health Q&A in traditional Chinese medicine.
Meanwhile, challenges have emerged in finding suitable application scenarios. Zhang noted that many companies are actually trying to do things that, in principle, are not suitable for large models, leading to poor results and even concerns about the reliability of large models.
“That’s why we often have customers map out their entire business process,” Zhang explained. “We start with the horizontal axis, showing what they do and the costs involved, and the vertical axis, showing the capabilities of large models. Then, we look for the areas within this two-dimensional matrix where large models are most suited. After identifying enough areas, our understanding of both the models and the business evolves, allowing us to draw a new business process map. Through this iterative process, we find the greatest common ground between businesses and large models and identify the most suitable collaboration scenarios.”
He illustrated this with an example from the pharmaceutical industry. When a new drug is launched and introduced to hospitals, reps need to answer questions from doctors about dosage, efficacy, and usage contraindications. In the past, reps often had to relay these questions to the medical affairs department, which would then spend days or even weeks reviewing experimental reports and research papers to ensure compliance before providing answers. However, with large models, the content delivered to the medical affairs department is already logically processed and systematized, similar to a standard operating procedure (SOP). It could take just 1–2 hours to confirm the accuracy of the content, which the reps can then relay to doctors.
As for pricing, it remains difficult to generalize about the price range that pharmaceutical companies or medical institutions are willing to pay for large model services. “You can spend a few thousand RMB to try it out, or spend hundreds of millions to implement it—it all depends on how you use the model,” the industry insider said.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Hu Xiangyun for 36Kr.