Peking University Cancer Hospital reduced average medical record writing time from about 3 minutes per record to under 20 seconds using a traceable AI-assisted documentation system. The hospital reports an efficiency improvement of more than 80% and a reduction of more than 30% in documentation work time. The system is in routine use across oncology medicine, surgery and radiotherapy departments.
The hospital faced heavy outpatient and inpatient documentation workloads. Manual electronic record entry was time-consuming and uneven in quality, while general-purpose language models raised concerns about hallucinations, compliance and response speed. The hospital worked with a third party to develop a system in which generated phrases can be checked against their sources.
Implementation proceeded through requirements and clinical standards, development and integration, limited clinical validation, hospital-wide deployment, and ongoing operational improvement. Pilot departments helped refine generation quality, responsiveness and usability. The broader rollout included physician training and adaptation to clinical workflows.
The technical stack includes Qwen-ASR speech recognition, Qwen-VL multimodal processing and OCR models. It runs on the hospital's internal GPU and CPU servers with Kubernetes and MongoDB. Information comes from patient pre-consultation input, recorded doctor-patient conversations, photographed medical documents, historical records and examination and laboratory results retrieved from the clinical data centre.
An extraction engine combines language models, traditional natural language processing and medical logic to identify disease labels and clinical events. Dedicated builders assemble individual record fields, including oncology histories involving staging, metastases and treatment. Generated passages retain source references and risk-level colour markings so clinicians can verify the underlying information.
The generation entry point is embedded in the existing EMR, allowing clinicians to insert generated text into the record without changing their usual interface. Outpatient records are generated in under 20 seconds, while inpatient documents take minutes. Data remains within the hospital network, with access controls, logging and encryption.
The source reports stable hospital-wide operation and subsequent deployment at other hospitals. Planned work includes deeper oncology capabilities, broader specialty coverage and a lighter version for primary care settings.