Abstract
This research, conducted at the Biomedical Big Data Center of West China Hospital, explores the construction and application of a breast cancer-specific database system based on the full data lifecycle. The study details the establishment of data standards, data fusion and governance, multi-modal knowledge graphs, and secure data sharing. By integrating structured and unstructured clinical data through NLP and standardized mapping, the system supports clinical practice, scientific research, and medical education, maximizing the value of medical big data.
The Challenge: Overcoming Data Silos
Medical research institutions frequently encounter "data silo" problems. Breast cancer treatment is inherently multi-disciplinary—involving surgery, chemotherapy, and radiotherapy—which generates vast amounts of heterogeneous data across disparate hospital systems like HIS, LIS, and PACS.
Researchers identified three primary obstacles:
- Data Volume: The scale of data exceeds the capacity for manual screening by clinicians.
- Standardization Gaps: Identical data elements (e.g., specific medications) are often coded differently across departments.
- Unstructured Complexity: Critical information is frequently trapped in unstructured text, such as pathology reports and electronic medical records, hindering accurate analysis.
Methods and Architecture
The researchers implemented a "full data lifecycle" approach organized into several technical layers, from raw data ingestion to front-end application.
- Metadata-Centered Governance: A master data standard was established, including a synonym database to map varied naming conventions to unified standards like ICD-10 and SEER.
- Automated ETL Pipeline: An Extract-Transform-Load platform performs incremental data extraction from hospital systems at regular intervals.
- NLP and Text Mining: Natural Language Processing (NLP) parses unstructured clinical text. A "back-annotation" process was used, where a model trained on 200 professional annotations assisted physicians in labeling a larger set of 1,000 samples.
- Multi-modal Knowledge Graph: The system constructs a node-network graph reflecting hierarchical relations between pathogenesis, symptoms, complications, and medications.
Dataset Scale and Structure
The system was built using a massive internal dataset from the Biomedical Big Data Center of West China Hospital, ensuring high granularity and clinical relevance.
- Scale: 7,730 patient cases registered since 2008.
- Structure: 20 database tables containing 444 distinct fields.
- Scope: Includes socio-demographic information, exposure/risk factors, inpatient/outpatient records, diagnostic imaging, and pathology results (ER, PR, HER2, Ki67 status).
Key Insights: AI and the "Holographic" Portrait
- AI as a Quality Gatekeeper: Beyond analysis, AI is utilized for data governance. Deep learning models automatically identify and filter out low-quality or irrelevant medical images before they enter the diagnostic database.
- The "Holographic" Portrait: The system aims to create a "holographic knowledge portrait" of the patient. This provides a multi-dimensional view of risk profiles and disease trends, assisting junior physicians and reducing misdiagnosis rates.
- Closing the Loop: Insights gained from the database are used to optimize existing clinical research processes, creating a sustainable development loop for hospital operations.
Results and Performance
- NLP Performance: The NLP extraction for electronic medical records achieved an accuracy of 80–85%, comparable to manual recognition by general physicians.
- Data Granularity: The cleaning process successfully mapped cases into a highly detailed structure, including 2.6 million outpatient order records and 8.2 million inpatient order details.
- Clinical Utility: The system currently supports a "Smart Patient Index" (EMPI) and clinical decision support tools utilized for teaching and research at West China Hospital.
Limitations and Future Directions
While NLP accuracy is high, high-stakes clinical applications still require expert review to ensure absolute reliability. Current implementation is centered on West China Hospital; future work will focus on expanding to a multi-center joint research network using Federated Learning and Multi-party Secure Computation to preserve privacy. Additionally, the team plans to develop "dynamic" knowledge graphs that automatically expand as new medical literature is ingested.
Summary updated