Concept-based electronic health records: opportunities and challenges

Shahram Ebadollahi, Anni Coden, Michael A. Tanenblatt, Shih-Fu Chang, Tanveer Fathima Syeda-Mahmood, Arnon AmirACM Multimedia 2006 · DOI

TL;DR

This paper proposes a "concept-based" approach to organizing electronic health records (EHRs), moving beyond document-level storage to a more intuitive, information-level structure. By linking multimedia data (images, videos, signals) and text to specific medical concepts, this paradigm aims to unlock decision-enabling information, improving healthcare efficiency and patient outcomes.

What Problem Does This Paper Address?

The current healthcare system is drowning in data but starving for actionable information. Terabytes of multimedia medical data are generated daily, but accessing relevant insights for timely decision-making is incredibly difficult. Existing Electronic Health Records (EHRs) aggregate data at the document level, meaning clinicians still have to sift through individual reports and images to piece together information about a specific condition or anatomical part. This fragmentation leads to inefficiencies, potential medical errors, and a missed opportunity to leverage the full potential of multimedia data for diagnosis, research, and patient education. The paper argues for a paradigm shift from document-centric to information-centric organization, focusing on the underlying medical concepts.

Methods

The core of the proposed solution is a concept-based multimedia medical record organization. This involves:

  • Concept Extraction Framework: Analytical engines (AEs) process various data sources (e.g., echocardiogram videos, pathology reports) to extract and identify "snippets of information" related to specific medical concepts. These AEs are guided by domain knowledge and ontologies.
  • Ontology-driven Organization: A hierarchical structure, often based on ontologies like the Unified Medical Language System (UMLS), defines medical concepts and their relationships. This provides the semantic backbone for organizing the extracted information.
  • Multimedia Content Structuring: This is a crucial enabling technology. It involves techniques to understand and tag elements within multimedia data, linking them to concepts. For images, this means segmentation and localization; for videos, it involves temporal segmentation and identifying key events; for text, it involves natural language processing to extract relevant phrases and semantics.
  • Concept-Based Health Record (CBHR): The output is a "growing attributed graph structure" where nodes represent concepts and links represent relationships. Each node is associated with attributes embodying manifestations of that concept across different data sources.
  • Coordinated Multimedia Summary (CMS): A user interface that dynamically generates navigable multimedia summaries based on user queries about specific concepts. It integrates snippets from various sources (videos, text, definitions) into a cohesive view.

Datasets

The paper does not explicitly list or introduce specific datasets for training or evaluation. Instead, it refers to the general types of data encountered in healthcare:

  • Medical Imaging: Magnetic resonance, functional magnetic resonance, ultrasound, computed tomography, positron emission tomography, microscopy, pathology images.
  • Video Data: Echocardiogram videos are specifically highlighted.
  • Signals: Electro-cardiogram (ECG), electro-encephalogram (EEG).
  • Textual Data: Lab reports, clinical notes, pathology reports, diagnosis reports.

The authors mention their prior work on echocardiogram videos and pathology reports as case studies, implying these were used in their research, but specific dataset details are not provided.

Key Insights and Ideas

  • The "Information-Poor" Healthcare Domain: Despite vast data, healthcare is "information-poor" because the data is not organized for easy access to decision-enabling insights.
  • Beyond Document-Level: The fundamental limitation of current EHRs is their document-centric nature. The true value lies in extracting and organizing information about concepts, not just storing the documents themselves.
  • Multimedia as a Key Enabler: Multimedia data (images, videos) are not just supplementary; they are critical for understanding patient conditions. Concept-based organization allows for the seamless integration of these diverse data types.
  • Ontologies as the Semantic Backbone: A well-defined ontology is essential for understanding the relationships between medical concepts and for guiding the extraction and organization of information.
  • Dynamic, User-Centric Summaries: The CMS concept offers a powerful way to present complex information tailored to specific user needs, allowing for interactive exploration of medical concepts.
  • Potential for Statistical Decision Support: By aggregating concept-manifestations across many patients, the system can enable statistically-guided decision support, allowing clinicians to compare their diagnoses with those of peers.

Key Results

The paper is primarily conceptual and proposes a framework. It does not present quantitative results from a large-scale implementation. However, it highlights the potential benefits and showcases work in two case studies:

  • Echocardiogram Video Structuring: The authors describe a system that automatically decomposes echo videos into temporal units (views, modes, cycles) and tags cardiac chambers. This work was used in the context of the PERSIVAL system for personalized multimedia summaries.
  • Cancer Pathology Report Structuring: MedTAS (Medical Text Analysis System) is presented as a tool that populates a cancer disease model by extracting information from pathology reports, aiming to structure unstructured text for better retrieval and analysis.

The key "result" is the conceptual framework and the demonstration of its feasibility through these case studies, suggesting that concept-based organization is achievable and beneficial.

Limitations and Future Work

The authors acknowledge several limitations and areas for future work:

  • Realization Challenges: Materializing concept-based health records requires overcoming significant hurdles, including defining a common, pluggable architecture for analytical engines and developing robust analytical engines.
  • Semi-Automated Analytics: For modalities where fully automated analytics are not yet feasible, semi-automated or human-in-the-loop approaches may be necessary.
  • Information Needs Mapping: Further investigation is needed on how to effectively map the information needs of different user types to the relevant portions of concept-based records.
  • Incremental Information Gain: Analyzing the incremental information gain of using multiple media types for a given concept is crucial for optimizing presentation.
  • Seamless Integration: The integration of this new paradigm into clinical settings and workflows must be seamless for users.
  • Future Work: The authors suggest exploring other multimedia data types (EKG, Cardiac CT, MRI) for decision support and further developing the multi-modal mining workbench for deeper analysis.

Summary updated

BibTeX
@inproceedings{ebadollahi2006concept,
  author = {Shahram Ebadollahi and Anni Coden and Michael A. Tanenblatt and Shih-Fu Chang and Tanveer Fathima Syeda-Mahmood and Arnon Amir},
  title = {{Concept-based electronic health records: opportunities and challenges}},
  booktitle = {ACM Multimedia},
  year = {2006},
  doi = {10.1145/1180639.1180859},
}
Figure