Unique Research · Interview · April 12, 2024
Background
On January 24, 2024, Unique Capital and partners held the 2024 AIGC Application Development Summit in Shanghai, with nearly 1,000 AI practitioners attending. This piece is edited from the guest keynote / roundtable segment.
Guests
- Dr. Xue Chong (薛翀) — Founder & CEO, Quanzhentong (全诊通)
- Qin Fei (秦飞) — Co-founder & CFO, Yizhe AI / 医者AI
- Qiao Minghui (乔明辉) — Founder & CEO, Yuemi Technology (悦米科技)
- Xue Yuanxin (薛元新) — Founder, Hengfang Health (横方健康)
- Moderator: Wu Shenliang (吴申亮) — Investment Advisor, Unique Capital
Roundtable
Quanzhentong — AI in care: from pre-visit to in-visit
Q: In healthcare, what successful cases or scenes of AI in in-visit and pre-visit stages are worth sharing?
Dr. Xue Chong, Founder & CEO, Quanzhentong: As a Western-medicine physician, I see AI and medicine as long intertwined. Since AI was coined at the 1953 Dartmouth conference, AI has played a major role in healthcare — especially imaging diagnosis and medical devices. After the classic-AI era we entered the foundation-model era; from last year, deep learning and AI ideas regained attention, giving us a chance to rethink and reapply these technologies.
After returning to China I wanted to build a physician-assistant product: the U.S. has physician assistants; China does not. U.S. PAs handle much peripheral physician work — charting, follow-up, and more. I kept asking whether China could use AI for those problems. For the past 8 years our company has focused on an AI product that helps doctors write medical records. When the foundation-model era arrived we adopted the tech promptly and rebuilt the product.
Care is usually split into pre-visit, in-visit, and post-visit. Pre-visit covers registration and appointments; in-visit covers work inside diagnostic decision-making; post-visit covers follow-up. Out-of-hospital patient management is another important concept. Within those buckets we mainly focus on AI assistance inside the diagnostic process.
On registration and appointments, Quanzhentong works on connecting and matching patients with hospitals. We built an AI product that can pop up after a patient registers, help the hospital and patient collect history materials, and generate a case summary — so when the patient arrives the doctor can review materials in advance and raise efficiency, especially in the in-visit stage.
In devices, classic AI — algorithms and models — has long been core to CT, MRI, and similar modalities. Our AI products have progressed further in imaging diagnosis, e.g. deep learning for pulmonary nodules and ultrasound analysis. On text, with foundation models we are deepening exploration. For clinical decision support we used to rely on knowledge graphs and rules; increasingly we use general AI to address out-of-hospital links as well.
Yizhe AI — out-of-hospital market: AI for health management
Q: Beyond the in-hospital market, out-of-hospital is also important. What is your view on concrete AI applications in the out-of-hospital market today?
Qin Fei, Co-founder & CFO, Yizhe AI: I previously worked in investing — first fund-of-funds, later mainly AI and new energy. Yizhe AI’s team is largely from Tsinghua, focused on out-of-hospital care, consumer healthcare, and health management. We use an MoE (mixture-of-experts) architecture that has already been recognized in places such as European MR. From September 2022 we began trying MoE with fine-tuned large models; in health management, MoE outperformed large models by a wide margin.
In out-of-hospital and consumer healthcare we focus on several key scenes. First: checkups — especially the private checkup market, where our shareholder Dajiankang Group (大健康集团) is an industry leader. We start from checkups with a focus on pulmonary-nodule screening. Current detection of pulmonary nodules has reached 54.3% — roughly one in two people — up from about 30% before the pandemic, with a clear rise during COVID. Given that lung cancer is among China’s most common fatal diseases, our strategy is early nodule identification to shrink the potential patient pool.
In actual hospital practice, nodules under 8 mm usually need no immediate treatment. We use AI models to analyze nodule etiology, simulate expert diagnostic reasoning, and deliver personalized plans. Data show 10% of patients see nodules shrink without treatment. Traditional hospitals often prescribe anti-inflammatories, effective in about 30% of patients; with our customized plans that share can rise to 70%. We aim to manage and intervene on checkup-found nodule patients so disease does not progress to surgery or needle biopsy.
That led us into insurance. At the core of insurance is family-doctor service for mid-to-high-end clients; family-doctor complaint rates remain high. We replicate a model spanning nutritionists, health managers, sports and rehab AI, plus disease-specific AI, to lift overall health-management outcomes — with meaningful gains on data metrics.
We are also entering dentistry and diabetes — all part of health management. Those are our main directions now.
Yuemi — AI optimizing service efficiency in patient management
Q: Beyond patient health management, patient management also involves operations and education, where many AIs are being tried. Can you share your view and experience?
Qiao Minghui, Founder & CEO, Yuemi Technology: I founded Yuemi and have been a healthcare entrepreneur for nearly a decade. We started in pharma — digital marketing and patient management — and with AI’s rise keep iterating; today we focus more on personalized service and AI applications.
We serve patients from private-domain ops through post-visit management, combining AI to find which scenes and key nodes we can help. One project underway: guests mentioned pulmonary nodules; lung cancer is China’s #1 cancer, while breast cancer is the #1 cancer for women and incidence is very high. Upstream of breast cancer are breast nodules. We are partnering with Danaher on an integrated management program for breast cancer and breast nodules involving three core roles: patient, doctor, and online medical assistant.
Serving patients, we focus on those three roles and multiple scenes for better experience. How to handle patients’ fragmented, diverse, personalized questions? We use an “82 principle”: 80% of questions answered by AI; of the remaining 20%, 16% by online medical assistants; the final 4% need specialist doctors. That split sharply raises efficiency and cuts labor cost. Breast-nodule and breast-cancer patients also have diverse needs: after checkup or screening, nodule grading changes and disease course progresses — so our AI tools provide knowledge education and interpretation.
Second role — doctors: in outpatient service more than 1/3 of a doctor’s time is not medical diagnosis but basic service and emotional support. Is there a better scene, and more efficient people and tools, for that mismatch? Helping doctors reduce burden matters; AI helps a lot here.
Third role — online customer service: they take some of doctors’ basic work, move patients from offline to online and from public to private domain for fuller patient management including mid-course follow-up and adverse-event management. We deeply integrate AI to raise online CS efficiency: previously one agent might serve 1,000 patients; now 3,000–5,000 — gains realized step by step against real business.
Hengfang Health — digital doctor–patient relations: smart science education and compliance
Q: On the patient-management block, can you share your experience and view?
Xue Yuanxin, Founder, Hengfang Health: I’m founder and CEO of Hengfang Health — a Huawei-lineage company; tech team from Huawei. We recently signed a strategic partnership with Huawei’s Pangu foundation model to build scene applications in healthcare. Our main focus is intelligent science-education projects, with patient education as the entry.
In hospital visits, patients often have only a few minutes with the doctor — tense doctor–patient relations, little sense of being served. Everyone here has been a patient. In that short window patients often cannot ask everything; many questions are missed; relations tighten. Doctors lack bandwidth — dozens of patients a day, mostly repetitive questions — creating information asymmetry. We want digital means to turn 60–70% of repetitive outpatient questions into science-education materials, easing pressure and improving relations.
Science education itself is subtle. Traditional medical popularization is already plentiful online (Tencent Medpedia, Baidu Health, etc.), but what users lack is not a medical library — it is getting the right, precise knowledge in natural language. Actively studying medicine is often anti-human: images and information can be uncomfortable; too much or too little learning also causes problems. So we propose “not too early, not too late; not too much, not too little” — the right amount of knowledge at the right time, so patients have something to hold onto when they temporarily need it.
AI can dialogue from patient need and state, and push foundation-model-generated knowledge points best fit for them — delivered via highly trusted doctor digital-human personas covering diet cautions, drug use, rehab exercise, and more. With Pangu we are already piloting with thousands of doctors nationwide, with good results.
Q: Does AI in healthcare face similar risks — or what challenges does the industry face today?
Xue Yuanxin: Healthcare, more than many industries, has a built-in tight hoop: compliance. Three layers. First, algorithm compliance: foundation models are powerful, but uncontrolled hallucination in care can create serious health risk — patients cannot judge truth. Second, data and privacy compliance — privacy protection in training and runtime. Third, ethics compliance: large firms driven by commercial interest may discriminate or select patients — also a compliance concern. Other challenges: if tech works too well, doctors may resist, fearing replacement. We need to solve a chain of challenges for compliant, sustainable medical tech.
Yuemi — AI from pharma and hospital angles: doctor–patient relations and pharma marketing
Q: Healthcare’s core principals are pharma companies and hospitals (doctors being part of hospitals). From those two principals’ angles, what are their attitudes and expectations toward AI?
Qiao Minghui: Yuemi’s model is mainly B2B; direct clients include drug and device companies. First, respect healthcare’s particularity. AI opportunities in smart CS, image/text generation, and video production exist elsewhere — and healthcare demand is also broad, because part of healthcare work lacks strong industry specificity. Inside healthcare’s particularity we also see interesting tracks: assisted diagnosis, smarter health management, and AI-enabled new-drug R&D — where strong companies and unicorns have already emerged.
Healthcare requires seriousness, professionalism, and compliance. When serving drug/device companies and hospitals, quality and safety bars are higher than FMCG or entertainment — think food-grade vs drug-grade; drug-grade is clearly stricter.
Data silos are also sharper than in other fields; fragmentation is worse. Broader AI use in healthcare needs a better big-data foundation. Zooming into pharma, devices, and hospitals — with marketing experience — several points matter.
First: AI will further substitute sales-rep roles. In China one medical rep serves 1.3 doctors; in Japan 5.2; in the U.S. 12. As a productivity tool AI has huge potential here to cut cost, raise efficiency, and push compliance.
Second: pharma and hospital-doctor demand will change traditional medical/academic promotion. Classic promotion centers on two scenes — conferences and department meetings. With AI we must rethink how to meet doctors’ academic needs in new ways.
Another scene is one-to-one visits — offline and online. Online visits used to mean “virtual reps” between human reps and AI, mostly call-center or systematized medical-info push. AI leaves large room to optimize visits. As human-rep headcount falls, quality and capability bars rise: reps must know medicine better and know AI better, to raise contact and influence efficiency with doctors.
Also: anti-corruption in healthcare runs through 2026; compliance will be a “dragon-slaying blade” overturning traditional commercial and academic promotion. AI has huge demand in reconstructing academic-promotion models.
Finally, in specialty-disease domains such as pulmonary and breast nodules, collaboration among clients may matter more than competition. Some breast-nodule patients progress to breast cancer — a funnel. Breast cancer has subtypes; AI and big data can promote collaboration and information reuse, not only competition.
So AI can reshape healthcare to meet drug/device companies’ and hospital doctors’ needs.
Doctor’s view — embracing new tech
Q: We just discussed from hospital or pharma angles; from the user-need angle, Dr. Xue, anything to add?
Dr. Xue Chong: Personally, as a doctor, my attitude to AI is positive. Excellent doctors tend to embrace new tech — the philosophy is solving complex medical problems with the simplest methods and helping patients with minimal harm. If AI can do that, doctors will welcome it.
From a management angle, healthcare is multi-headed, tightly regulated, and closed — many compliance requirements. Management must consider business model: whether patients pay, which ties to whether a medical-device certificate is needed.
In China AI products can obtain medical-device certificates, but the business model must match. If the AI product does not charge patients and is provided by the hospital or as infrastructure, some compliance issues can be avoided. Overall, doctors want technologies that help patients with minimal side effects — so the newest, most advanced tech often lands first in large hospitals.
Yizhe AI — outlook: cutting cost and personal health twins
Q: What are your expectations for AI in healthcare this year?
Qin Fei: Two scenes. First: AI–healthcare combination is an obvious trend, but AI — especially foundation-model inference — is very expensive. Cutting cost is therefore critical. How?
Example: we are building an edge-model path — essentially MoE. An MoE has a base scheduling model underneath with many small models plugged on top. The scheduler knows each small model’s capability range and dispatches different small models to collaborate on a task. Training cost falls sharply: vs putting the same data into a single model with tens of billions of parameters, a scheduler-plus-many-small-models edge model’s training cost drops exponentially.
Second: we deploy some small models on the edge — e.g. on each person’s phone — so real inference runs on-device. Phone compute is already enough; that greatly cuts centralized-server cost while raising efficiency and response speed. That is one cost-down direction for foundation models in healthcare.
Another direction: combine AI with health management. We see a personal health twin as important — and are working to land it. From MoE we develop a small model per person. In dialogue with AI, checkup reports, clinical conclusions, and daily smart-device data all feed that personal small model, forming a personal health digital twin that gradually knows the person’s health better than they do.
Example: next time you need a specialty visit, the twin acts proactively — finds top doctors in the field or the nearest hospital, helps book, and tells you where to go and how efficient the booking is — saving your own search, asking, and registration time.
The twin can also predict disease to some degree. Health management then intervenes early to prevent onset. We see that as another major direction and trend.
Looking ahead — multimodal foundation models and agents
Q: Dr. Xue, how do you see future AI–healthcare trends?
Dr. Xue Chong: Some forecasts for future healthcare are uncertain. Medical products, like other industries, mix labor output and decision-oriented cognitive output. Internists, for example, make a series of judgments and decisions and gather information through doctor–patient dialogue. Bill Gates, Mustafa Suleyman, and others believe large models will open many opportunities in healthcare. Future large models may re-process medical data — e.g. detect pulmonary nodules and generate corresponding reports. Such multimodal combinations can give richer information for more accurate diagnosis and treatment decisions.
Medical data is inherently multimodal; future large models can re-process it. Software inside devices can also be upgraded via large models; some are even trying large models to control robots for perception, decision, learning, and output — among the best practices in healthcare. Example: a LASIK laser for myopia can measure eye distance itself, plan corneal ablation, then let the machine execute — that autonomous agent is an Agent.
In the future, combining machine vision with these models, healthcare will see all kinds of progress that make auto-completion of many tasks possible. That is roughly the future I can foresee — though I don’t know how long it will take to arrive.
2 CN images (event / guests); attach on ingest.
— END —