The Stanford Medicine data science ecosystem for clinical and translational research

Alison Callahan, Euan A. Ashley, Somalee Datta, Priyamvada Desai, Todd Ferris, Jason Fries, Michael Halaas, Curtis P. Langlotz, Sean Mackey, Jose Posada, Michael A. Pfeffer, Nigam H. ShahJAMIA 2023 · DOI

TL;DR

Stanford Medicine has built a comprehensive data science ecosystem (SDSR) to facilitate clinical and translational research by integrating diverse patient data sources into a unified repository (STARR). This ecosystem provides researchers with the infrastructure, tools, and services needed to access, analyze, and derive insights from de-identified patient data, paving the way for a learning health system.

What Problem Does This Paper Address?

Academic medical centers (AMCs) are increasingly recognizing the need for robust data science capabilities to advance clinical and translational research. However, managing the vast and disparate data generated by modern healthcare IT systems presents a significant challenge. These systems capture data with varying granularity and for different purposes, making it difficult to create a comprehensive view of a patient's journey. The problem is to build an integrated infrastructure that can effectively ingest, process, de-identify, and provide secure access to this data for research purposes, while also supporting advanced analytics and machine learning applications. This is crucial for building learning health systems that can continuously improve patient care through data-driven insights.

Methods

The core of the Stanford Medicine Data Science Resources (SDSR) is the Stanford Medicine Research Data Repository (STARR), which utilizes the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM). This standardized model facilitates interoperability across different data sources and research centers.

Key methodological components include:

  • Data Ingestion and Processing: A robust Extract-Transform-Load (ETL) pipeline is used to migrate data from separate Epic instances (Stanford Health Care and Stanford Children's Health) to Google Cloud Platform (GCP) BigQuery. This process involves data de-identification and the creation of persistent, de-identified "person identifiers" to link records across datasets.
  • De-identification: A two-step process is employed for de-identifying clinical text, using a hybrid NLP approach (TiDE) combined with pattern-matching heuristics, followed by a "hiding in plain sight" (HIPS) method to replace identifiers with surrogates. For clinical images, a distributed software application operates on-demand, using custom filtering rules based on the DICOM standard.
  • Secure Computing Infrastructure: Researchers access de-identified data through Nero, a secure data science platform managed by the Stanford Research Computing Center (SRCC). Nero offers self-service access to compute environments with tools like Jupyter notebooks, Python, and R, as well as OHDSI tools.
  • Data Linkage: The system creates persistent "person identifiers" that allow for automatic linkage of records across STARR datasets. This enables a comprehensive view of a patient's history.
  • Entity Extraction: A natural language processing (NLP) pipeline is used to extract clinical terms from provider notes, creating a dictionary of terms that can be used for advanced phenotyping.

Datasets

The paper describes several key datasets within the STARR:

  • STARR OMOP: Contains de-identified clinical data from over 3.4 million patients, including demographics, diagnoses, medications, lab results, and clinical notes.
  • STARR Radio: Houses radiology and cardiology imaging data from Stanford's PACS, comprising over 5.7 million studies and 1 billion DICOM records.
  • STARR Wave: Includes pediatric bedside monitoring data (heart rate, blood pressure, etc.) and alerts, linkable to EMR and radiology data.

The paper also mentions the public release of 19 imaging datasets, compiled from Stanford Health Care patient records, for the purpose of developing and validating machine learning applications.

Key Insights and Ideas

  • The Power of Standardization: Adopting the OMOP CDM is crucial for enabling interoperability and network studies, allowing researchers to leverage community-developed tools and collaborate with other institutions.
  • Balancing Access and Security: The SDSR demonstrates a successful model for providing broad access to de-identified research data while maintaining robust security and privacy controls through secure computing environments and data use agreements.
  • The Importance of a Comprehensive Ecosystem: Beyond just data storage, the SDSR integrates infrastructure, processing pipelines, de-identification services, analytical tools, and user support (training, office hours) to create a holistic research environment.
  • The "Honest Broker" Model: The use of "honest broker" personnel within Technology and Digital Solutions (TDS) is key to managing the de-identification and linkage processes, ensuring data privacy.
  • Iterative Development: The SDSR evolved from an earlier platform (STRIDE) and continues to adapt, highlighting the dynamic nature of building and maintaining such complex systems.

Key Results

  • The STARR repository integrates data from multiple sources, including EMRs, clinical images, text, and bedside monitoring data.
  • The SDSR supports over 120 faculty research groups, more than 800 cloud compute users, and over 300 self-serve de-identified STARR data users.
  • Since March 2020, the number of STARR data users has increased by 86%, and the number of principal investigators using STARR data and services by 70%.
  • The system has enabled the public release of a large number of imaging datasets, contributing to the broader research community.
  • The SDSR ecosystem supports diverse research activities, including hackathons, multi-institutional collaborations, and health system implementations.

Limitations and Future Work

  • Data Omissions due to OMOP CDM: The paper acknowledges that conforming to the OMOP CDM requires omitting certain data elements from the original EMRs (e.g., detailed medication instructions), which can limit the scope of analyses.
  • Translation of Research to Practice: The paper notes that translating research innovations (like machine learning models) into operational improvements remains a bottleneck, leading to the formation of a new data science team within Stanford Health Care to streamline this process.
  • Financial Sustainability: Ensuring sufficient financial support for the SDSR infrastructure and personnel is an ongoing challenge, requiring a mix of institutional funding, grants, and cost recovery models.
  • Balancing Tool Development and User Training: Striking the right balance between developing new tools and providing effective user training is critical for maximizing researcher adoption and minimizing the need for intensive concierge services.
  • Future Work: The authors aim to continue expanding the SDSR by incorporating more diverse data sources and developing tools that support a wider range of downstream use cases. They also hope their experience serves as a valuable case study for other AMCs building similar ecosystems.

Summary updated

Abstract

Our experience maintaining the SDSR offers a case study for academic medical centers developing data science and research informatics infrastructure.
BibTeX
@inproceedings{callahan2023stanford,
  author = {Alison Callahan and Euan A. Ashley and Somalee Datta and Priyamvada Desai and Todd Ferris and Jason Fries and Michael Halaas and Curtis P. Langlotz and Sean Mackey and Jose Posada and Michael A. Pfeffer and Nigam H. Shah},
  title = {{The Stanford Medicine data science ecosystem for clinical and translational research}},
  booktitle = {JAMIA},
  year = {2023},
  doi = {10.1093/jamiaopen/ooad054},
}
Figure