{
  "abstract": "Background The Quality Improvement (QI) team have been working on an internal project for identifying patients who are at higher risk of fractures within the Trust. Based on their initial analysis, radiology reports were considered as the best source of data for determining the rate of inpatient fractures. A significant amount of manual effort is required in a timely manner from an expertise with domain knowledge for understanding of the reports. Given most reports do not contain any fractures detected, this becomes a minority labelling issue that may introduce annotation inconsistencies which can require further investigation.Solution The DRIVE team in collaboration with the QI team have developed an automated Agentic AI system that can seamlessly process unstructured text information in radiology reports for identifying potential pathological fractures, successfully detecting new types of pathological fractures and recognising similar fractures of the same patient over the time which is critical to identify whether it is a new fracture or not. By utilising an LLM-based Agentic system, it can handle information about rare diseases. Our proposed Agentic system uses an open-source LLM, named MedGemma, which has been specifically trained on medical text and comprehension for performing the following actions (1) analyses a report for fracture, (2) collects evidence present in the report, (3) generates a rationale based on evidence and (4) flags if a patient has a pathological fracture. The outputs are used by the QI team and the consultants to make a final decision. ( Figure 1)Conclusion Our internal validation of the Agentic system as a baseline has identified a significant performance improvement based on the user feedback . This has reduced the process from 2-3 days of manual work to a few minutes and allowed team to demonstrate a process for collecting more accurate number of fractures identified by using radiology reports.Abstract 139 Figure 1",
  "authors": [
    {
      "affiliations": [
        "Great Ormond Street Hospital for Children NHS Foundation Trust, UK"
      ],
      "name": "Sebin Sebu"
    },
    {
      "affiliations": [
        "Great Ormond Street Hospital for Children NHS Foundation Trust, UK"
      ],
      "name": "Nuwanthi Yapa Mahathanthila"
    },
    {
      "affiliations": [
        "Great Ormond Street Hospital for Children NHS Foundation Trust, UK"
      ],
      "name": "Alexander Chesover"
    },
    {
      "affiliations": [
        "Great Ormond Street Hospital for Children NHS Foundation Trust, UK"
      ],
      "name": "Pavithra Rajendran"
    },
    {
      "affiliations": [
        "Great Ormond Street Hospital for Children NHS Foundation Trust, UK"
      ],
      "name": "Lucy Turriff"
    },
    {
      "affiliations": [
        "Great Ormond Street Hospital for Children NHS Foundation Trust, UK"
      ],
      "name": "Shiren Patel"
    }
  ],
  "title": "139 Agentic AI for automated classification of paediatric radiology reports",
  "uid": "0ffa5a90-e319-5654-971a-7ff65fffef42"
}