|
Non-Disclosure Agreement: NO
Intellectual Property: YES
Physical Prototype or On-Campus Equipment: NO
Images and Additional Links (if provided)
Hospitals collect patient safety event reports that contain valuable information about errors, near misses, contributing factors, and opportunities for improvement. However, these reports are often reviewed through labor-intensive manual workflows, and it can be difficult to identify related events or recognize themes that emerge across many reports over time.
The goal of this project is to design and build an integrated prototype platform that uses artificial intelligence and modern software engineering to support the review and analysis of hospital safety events. This is not solely an artificial intelligence modeling project. Students will design a broader software ecosystem that connects data ingestion, artificial intelligence models, secure data handling, storage, workflow management, and an intuitive user interface.
The prototype should ingest narrative safety reports and generate structured information that can assist with initial triage and organization. It should identify and link similar historical reports so users can easily explore related cases, recurring contributing factors, and potential patterns. The platform should also group reports into broader themes and display how those themes change over time.
A major emphasis will be placed on user experience. Students should develop an interface that makes safety events discoverable, allows users to move easily between related cases, explains why cases were linked, and presents trends in a clear and actionable manner. Creativity in the interface, visualizations, search functions, and methods of exploring relationships among events is encouraged.
The project should establish a privacy-conscious technical approach for developing, training, testing, and evaluating artificial intelligence components in a healthcare environment. The sponsor plans to provide the student team with fully synthetic safety-event data generated locally from the structure and general characteristics of real institutional event reports. The synthetic dataset will be designed to support realistic development and testing without exposing identifiable patient information or confidential event narratives.
The sponsor also maintains access to the underlying institutional safety-event data within the secure hospital environment. Students should therefore design the software and models so they can be packaged in a portable, containerized format. The sponsor can run the container internally against institutional data and return aggregate performance results, error examples, and other non-identifiable feedback to support iterative evaluation. Institutional data will remain within the sponsor’s environment and will not be transferred to the student team.
The proposed architecture should also address access control, data minimization, audit logging, separation of development and evaluation datasets, model versioning, reproducibility, and protection of sensitive information.
An additional advanced objective is to explore a prototype method for reviewing clinical records to identify possible safety events that were not submitted through the voluntary reporting system. The system may generate draft MIDAS-style event summaries for human review and inclusion in broader trend analyses. This component should focus on feasibility, architecture, representative workflows, and proof-of-concept methods using public, de-identified, or synthetic clinical data.
Expected deliverables include the following: • An integrated prototype platform with a functional front-end and back-end for ingesting safety reports, reviewing individual events, navigating linked cases, and exploring themes and trends over time. • Artificial intelligence models or model interfaces for report triage, structured information extraction, similar-case retrieval, and thematic grouping. The sponsor has developed preliminary models to address these functions, which may be used as baselines. Students will be encouraged to evaluate alternative approaches and refine or replace components when they can demonstrate improved accuracy, efficiency, interpretability, or usability. • A documented and reproducible data and system architecture, including pipelines for ingesting and validating sponsor-provided synthetic data, privacy-conscious development practices, access controls, auditability, and separation of development and evaluation data. • A containerized version of the models and supporting software that the sponsor can run within its secure institutional environment to evaluate performance using real safety-event data without transferring those data to the student team. • An evaluation framework covering model performance, retrieval relevance, usability, reliability, interpretability, failure modes, and appropriate human review. • Complete project documentation and handoff materials, including source code, system diagrams, setup and testing instructions, recommendations for future development, a final technical report, a demonstration, a poster, and a one-page summary.
The following is a "stretch goal" that we hope the team can achieve: • A feasibility assessment and, if achievable within the project scope, a proof-of-concept method for identifying potentially unreported safety events from clinical records and generating draft safety-event summaries for human review.
The overall emphasis should be on integrating these components into a coherent, creative, and usable end-to-end prototype rather than simply reproducing the sponsor’s existing preliminary models. |