Implementing a system for the real-time risk assessment of patients considered for intensive care

Simarjot S. Dahella, James S. Briggs, Paul Coombes, Nazli FarajiDavar, Paul Meredith, Timothy Bonnici, Julie L. Darbyshire, Peter J. WatkinsonBmc Medical Informatics Decis. mak 2020 · DOI

TL;DR

This paper introduces HAVEN, a novel real-time patient deterioration monitoring system designed to identify individuals at risk of requiring intensive care unit (ICU) admission. By leveraging innovative technologies like Portable Format for Analytics (PFA) and OSGi, HAVEN offers a configurable and dynamically updatable solution for early intervention, aiming to improve patient outcomes and reduce unplanned ICU admissions.

What Problem Does This Paper Address?

The paper addresses the critical issue of delayed identification of deteriorating patients in hospitals, which often leads to unplanned ICU admissions and poorer patient outcomes. Unplanned ICU admissions constitute a significant portion of total ICU admissions, and delays in recognizing the need for higher care are common, contributing to high short-term mortality. Existing systems often rely on static predictive models that are difficult to update, leading to suboptimal performance. The authors aim to create a system that not only identifies at-risk patients in real-time but also facilitates timely intervention by clinical teams.

Methods

The HAVEN system is built using a modular, web-based architecture. Key technologies include:

  • Java and Scala: For programming language flexibility and interoperability.
  • OSGi (Open Services Gateway initiative) Framework: This provides a modular system and service platform, allowing components (bundles) to be dynamically installed, started, stopped, and uninstalled without system reboots. This is crucial for enabling easy updates and maintenance.
  • Portable Format for Analytics (PFA): This is a new standard for describing statistical models in a JSON format. It allows for the definition of predictive models in a way that is easily configurable and executable, enabling the analytics team to provide models to the deployment team without requiring extensive engineering effort.
  • Hadrian Library: Used to execute PFA models.
  • MySQL 5.6: Serves as the aggregated system database for storing patient data.
  • ReactJS: Used for the client-side implementation of the user interface (UI), enabling dynamic and interactive visualizations.
  • Spring-boot Framework: Used for the server-side implementation of the UI.

The system's architecture is conceptually divided into layers, with data aggregation handled by the HAVEN Integration Engine (HIE), which processes data from various hospital systems (PAS, SEND, pathology). The HAVEN Scoring Engine (HSE) then uses the PFA-based model to compute the patient's risk score.

Datasets

The paper describes the system's deployment in a research project at Oxford University Hospitals NHS Foundation Trust. The system has processed data from:

  • Over 164,000 vital signs observations.
  • Over 68,000 laboratory results.
  • Data for over 12,500 hospital inpatients.

The data is sourced in real-time from existing hospital systems, including the Patient Administration System (PAS) for Admissions, Discharges, and Transfers (ADT), the Vital Signs collection system (SEND), and the hospital's pathology laboratory system. Messages are received in the HL7 version 2 format. No new benchmarks were introduced; rather, the system integrates with existing data sources.

Key Insights and Ideas

  • Dynamic Configurability is Key: The paper strongly emphasizes the importance of a system where the predictive model can be easily updated and reconfigured. This is a significant departure from traditional static models and is enabled by PFA.
  • PFA for Model Agnosticism: PFA's flexibility in defining arbitrary function compositions makes it ideal for statistical models, allowing for complex models to be implemented and easily swapped out.
  • OSGi for System Agility: The use of OSGi is highlighted as a major benefit, enabling a micro-services approach that reduces complexity, enhances scalability, and allows for dynamic updates without downtime. This is a novel application of OSGi in healthcare software.
  • User-Centric Design: The UI development involved close collaboration with clinical teams, using human factors methods like Cognitive Task Analysis and Card Sorting to ensure the interface is intuitive and meets user needs.
  • Separation of Concerns: The system's architecture adheres to this principle, with distinct components for different functionalities, making it more robust and maintainable.

Key Results

  • The HAVEN system has been successfully deployed and has processed a substantial volume of patient data (>164,000 vital signs, >68,000 lab results for >12,500 patients).
  • The system has demonstrated its capability to process large amounts of data effectively.
  • The final UI version was evaluated and found to perform well, affording accurate understanding with low cognitive workload for users.
  • No clinical decisions are currently being made based on the prototype system's output; evaluations are being conducted using retrospective data.

Limitations and Future Work

  • Prototype Stage: The system is currently a prototype and is being evaluated. It is acknowledged that it will require considerable evolution before widespread adoption.
  • Future Enhancements: Further system enhancements are planned to handle new data sources and additional management screens.
  • Clinical Decision Support: While the system identifies at-risk patients, it does not currently make clinical decisions. The focus is on providing information for clinician review and response.
  • Proprietary License: The system has a proprietary license, with restrictions on use by non-academics, requiring a license.

Summary updated

BibTeX
@article{dahella2020implementing,
  author = {Simarjot S. Dahella and James S. Briggs and Paul Coombes and Nazli FarajiDavar and Paul Meredith and Timothy Bonnici and Julie L. Darbyshire and Peter J. Watkinson},
  title = {{Implementing a system for the real-time risk assessment of patients considered for intensive care}},
  journal = {Bmc Medical Informatics Decis. mak},
  year = {2020},
  doi = {10.1186/s12911-020-01176-0},
}
Figure