Smart summaries of medical records

Learning Universal Patient Representations with Hierarchical Transformers

NIH-funded research Boston Children's Hospital · NIH-11144550

This project creates computer tools that turn long medical record texts, images, and codes into clear summaries to help doctors and researchers.

Quick facts

Grant typeR01 grant
Study typeNIH-funded research
Funding institutionBoston Children's Hospital NIH-funded
Lab location1 site (Boston, United States)
Project IDNIH-11144550 on NIH RePORTER

What this research studies

Researchers will use electronic health records—clinical notes, reports, images, and coded data from public and hospital sources—to teach new AI models how to summarize and combine these different data types. They will train smaller sentence- and paragraph-level encoders using a method called extreme distillation, then stack those into hierarchical transformer models to handle much longer notes. The team will experiment with different pre-training tasks and architectures and then merge text-based representations with structured data and images. Finally, they will apply these combined representations to clinical classification tasks relevant to emergency and chest-related care to find approaches that produce the most useful outputs.

Who could benefit from this research

Good fit: Patients whose electronic health records (clinical notes, reports, or images) are in participating hospitals or public datasets—especially those with complex or lengthy records—are the kinds of records this work uses.

Not a fit: People without electronic medical records, with very sparse records, or whose rare conditions are not represented in the training data may not see benefit from these models.

Why it matters

Potential benefit: If successful, this could give clinicians faster, clearer overviews of your health record to support quicker and more accurate decisions.

How similar studies have performed: Transformer-based language models applied to clinical notes have shown promising improvements in prediction and summary tasks, but combining hierarchical long-text transformers with imaging and structured data at this scale is still relatively new.

Where this research is happening

Boston, United States

Researchers

About this research

  1. This is an active NIH-funded research project — typically early-stage science, not a clinical trial accepting patient enrollment.
  2. Some NIH-funded labs run parallel clinical studies or seek volunteers for related work. To check, contact the principal investigator or institution listed above.
  3. For full project details, budget, and progress reports, visit the official NIH RePORTER page below.
Last reviewed 2026-06-13 by the Find a Trial editorial team. Information on this page is for educational purposes and is not medical advice. Always consult qualified healthcare professionals about clinical trial participation.