Page 49 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

support throughout the medical process, including triage, insurance consultation, and
report interpretation. For residents with general health needs, AI health assistant can offer
daily consultations on exercise, diet, mental health, sleep, and more.
• For Doctors: The solution offers doctors an intelligent assistant to aid in clinical care
and administrative tasks. The solution’s value lies in optimizing doctors' current "70% 4.1-Healthcare
administrative work + 30% clinical work" time allocation. AI handles tasks it excels at—such
as standardizing information and streamlining processes—freeing doctors to focus their
expertise on complex case analysis and doctor-patient communication, thereby release
more medical resources.
• For underserved areas: This case collaborates with top-tier medical teams to input the
clinical experience of national renowned doctors and departments into the model, creating
doctor agents or specialty-specific agents. Ordinary patients can access medical advices
without visiting hospitals. For complex cases or when in-person visits are necessary, the
intelligent agent can recommend nearby primary healthcare facilities, and local doctors
can combine the agent’s symptom collection and insights to deliver diagnoses, improving
both the efficiency and quality and also mitigating regional disparities.
All the above-mentioned solutions are built on the underlying capabilities of Ant Group's
medical large model. We will elaborate the technological innovations within the model in the
following section.

Technological Innovations
• High-Quality Medical Knowledge Graph and Professional Annotation Process:

o High-Quality Medical Data: To tackle the challenges of scarce high-quality medical
data, the project collaborated with hundreds of professional medical teams to build
a data system containing hundreds of billions of tokens of professional medical texts,
hundreds of millions of multimodal images and texts, and tens of millions of high-
quality medical knowledge graph entries, covering the full pre-, during-, and post-
diagnosis cycle with multimodal data formats (text, image, audio, video).
o Professional Annotation Process: Pioneered a fully automated annotation system
combining large model distillation technology and deep involvement from specialist
doctors, improving annotation efficiency by 50%. The RJUA series medical datasets
released with Shanghai Renji Hospital cover 85% of disease types, 97% of patient
demographics, and 96% of high-frequency lab items, setting an industry benchmark.
The project's data annotation and evaluation system also create a feedback loop to
continuously enhance data quality.

• Multimodal Interaction Accuracy Exceeding 90%, Achieving Industry-Leading Medical
Expertise:

o Multimodal Interaction: Built on Ant Group's self-developed medical large language
model and training optimization algorithms, the system supports multimodal data
exchange including images (e.g., medical imaging with resolution, timestamp, device
info, anonymized patient IDs, annotations), audio (e.g., symptom descriptions, doctor-
patient conversations), and video (e.g., rehabilitation training, remote consultations).
o Model Accuracy Over 90%: Using KG-CoT (Knowledge Graph Chain of Thought)
and PRM-MCTS (Process Reward Models and Monte Carlo Tree Sear) technologies,
the system achieves real-time optimization of diagnostic pathways and 98% tool
invocation accuracy. A Unified-Alignment Post-Training framework deeply integrates
reinforcement learning and instruction fine-tuning, achieving 93% diagnostic accuracy
and over 95% personalized service satisfaction.
o Medical Professionalism: The model surpasses GPT-4 in Chinese and English medical
evaluations, ranking first on PromptCBLUE-A and the Chinese medical benchmark
MedBench.

13