D24AC00379

Manual chart curation is slow, laborious, and expensive, which affects many facets of health research and clinical practice.

Collecting and mapping clinical data from electronic health records is critical for clinical and research services, including everything from doctors preparing to see a new referral patient to researchers assembling an international data sharing registry.

In response to the sociotechnical system innovation and health ecosystem integration interest areas of the ARPA-H Resilient Systems Office Office Wide Innovative Solutions Opening (ISO), we propose to develop and distribute the Democratized AI Guided Chart Abstraction Platform (DAGCAP).

DAGCAP is a modular, interoperable, comprehensive end-to-end workflow aimed at decreasing the chart curation effort by 90% while maintaining human-level accuracy.

Users will be able to precisely define data variables using a guided interactive process, and these variables will be mapped to commonly used data models as well as those defined or imported by users.

DAGCAP will revolutionize information extraction across healthcare and biomedical sciences, democratizing accurate chart extraction for all types of healthcare professionals.

Chart curation is a major bottleneck for oncology clinical practice and research.

Existing automated approaches are insufficient to address this bottleneck, as important patient data are stored as natural language rather than structured data.

A streamlined and automated curation system that mitigates these challenges and increases the quality of curation would drastically expand research, clinical, and operational capabilities across oncology care and research.

Large language models have potential to address insufficiencies of existing approaches, but correctly formulating the LLM inputs, prompts, and retrieval methods is critical.

Advances in large language models (LLMs) have created tremendous opportunities, but naive application of LLMs directly to chart abstraction produces inadequate results.

This project will address the calls for user-centric digital health tools that improve health outcomes and for novel ways to collect, integrate, analyze, and present health data by developing and releasing a prototype of the Democratized AI Guided Chart Abstraction Platform (DAGCAP) for semi-automated, machine-assisted semantic data curation.

DAGCAP allows any researcher, operational staff, or clinician to define a data dictionary, load a set of charts (directly from the medical record system or as flat files), and extract data elements accurately with minimal manual intervention.

The quantitative objective is to reduce the time for clinical chart abstraction by 90% (i.e., chart reviews that normally take an hour would take six minutes) while maintaining human-level accuracy.

Our proposed architecture will democratize clinical data curation and access to state-of-the-art clinical advancements.

The DAGCAP architecture will facilitate system scalability and sustainability in many user environments, increasing the ability of the cancer data ecosystem to incorporate data from low-resource environments, including those that serve marginalized patient populations.

DAGCAP will be accessible to all types of healthcare professionals across the full range of resource settings.

The modular design proposed will make our system scalable to additional data types, external datasets, and disease domains.

Vanderbilt University Medical Center

SUPPORT GROUND BREAKING RESEARCH IN SOCIOTECHNICAL SYSTEM INNOVATION HEALTH ECOSYSTEM INTEGRATION AND ADAPTIVE AND ANTIFRAGILE SOLUTIONS.

93.384 - ADVANCED RESEARCH PROJECTS AGENCY for HEALTH (ARPA-H)

Interior Business Center (IBC) [DOI]

Tennessee United States

State-Wide

D-AQD-FA-24-021

Amendment Since initial award the End Date has been extended from 08/31/26 to 08/31/27 and the total obligations have increased 101% from $1,981,115 to $3,978,849.