West China Hospital & Runda Medical

The Challenge

West China Hospital, affiliated with Sichuan University, is one of the largest and most prestigious medical centers in Asia, serving millions of patients annually. Physicians were spending up to 40% of their consultation time on administrative documentation — writing detailed patient records, transcribing symptoms, and logging treatment plans.

This administrative burden not only reduced the number of patients a doctor could see per day, but also led to inconsistencies in record quality, particularly during high-volume outpatient clinics where physicians might see 80+ patients in a single shift.

The Solution

In partnership with Runda Medical (a specialist health-tech company) and Huawei, the hospital deployed an LLM-based system running on Ascend inference servers that listens to doctor-patient conversations in real-time and generates structured electronic health records (EHR) automatically.

The system works in three stages:

• Speech Recognition: A specialized medical ASR model captures the conversation, handling medical terminology, dialect variations, and multi-speaker scenarios
• Semantic Understanding: A fine-tuned LLM on Ascend NPUs extracts symptoms, diagnoses, prescriptions, and follow-up recommendations from the transcript
• Record Generation: A structured EHR document is generated within ~1 second and presented to the physician for review and sign-off

"The system doesn't replace physician judgment — it eliminates the clerical overhead so doctors can focus entirely on clinical decision-making."

Technology Stack

Why On-Premise Matters Here

Patient health data is among the most sensitive information categories under China's Personal Information Protection Law (PIPL) — the Chinese equivalent of GDPR. Sending voice recordings and medical records to a public cloud API would create unacceptable compliance risks.

By running the entire pipeline on local Ascend hardware within the hospital data center, all patient data stays within institutional boundaries. Zero data ever leaves the premises.

Results & Impact

• Record generated in ~1 second — compared to 3-5 minutes of manual typing
• 60% reduction in physician admin time, enabling more patients per shift
• 95% accuracy rate on structured record fields (symptoms, diagnoses, medications)
• Full data sovereignty — no patient information leaves the hospital network
• Scalable architecture — system has been replicated across multiple hospital departments

Key Takeaway

Healthcare represents the ideal use case for on-premise AI: extreme data sensitivity combined with high-volume, latency-critical workloads. The Ascend inference stack delivers the performance needed for real-time NLP without any compromise on patient privacy. This model is directly transferable to European healthcare settings operating under GDPR.

📎 Source: Huawei Connect 2025, Runda Medical partnership announcement