Research Impact at Durham University: Applying AI to improve healthcare communication

Main ideas, concepts, and lessons

• Problem in healthcare communication (discharge letters)

  • Northern Care Alliance NHS Foundation Trust (NCA) manages four hospitals across Greater Manchester.
  • The trust produces 150,000–200,000 discharge letters per year.
  • It is almost impossible to comprehensively evaluate how effective discharge letter writing is.
  • Discharge letters are crucial for patient safety because they are the only information GPs receive when patients go home.
  • Discharge letters contain what community teams need for ongoing care.
  • Quality review is currently manual (usually by clinicians) and is time-consuming:
    • ~15–20 minutes per discharge letter

• Proposed solution: AI automation of discharge-letter audits

  • NCA partnered with the University of Stirling to automate the review/audit of discharge letters.
  • Motivation:
    • Previous work at NCA on A&E admission and discharge led the team to focus on what happens after discharge.
    • Findings referenced:
      • ~20% of adverse events occur after discharge
      • ~60% of those are due to poor communication
  • With AI language models, the goal is to scale up auditing from small efforts to every discharge letter, including communication content that occurs within the hospital.

• Interdisciplinary development approach

  • The work is supported by funding from the Impact Acceleration Account.
  • The project team is interdisciplinary, involving:
    • Clinicians
    • Computer scientists
    • Patients
    • Quality improvement teams from the hospital
  • Early meetings were used to align everyone on common goals, then the team integrated:
    • clinical expertise (what to assess / achieve)
    • computer science expertise (how to implement the AI)

• How the AI audit is intended to work

  • The team uses large language models to:
    • identify patterns in discharge-letter data
  • The AI outputs are then:
    • checked/assessed against a clinician’s judgment (human review of the automated outputs)
  • This human-check step supports quality control of the AI audit results.

• Reported outcomes and impact

  • Reported performance:
    • The AI can review every discharge letter generated across all hospitals over a year in ~15 minutes—matching the approximate time a human would need to review one letter.
    • The model’s quality is described as as good as a human, and possibly better, because:
      • it can detect themes/patterns humans may miss due to limited exposure.
  • Further benefits attributed to the project:
    • A postdoc gained/resecured a fellowship
    • A spin-out company was created from the project
    • A large NIH R application was submitted

• Broader lesson / takeaway

  • The project is presented as an example of responsible and safe AI use in NHS settings to:
    • improve patient safety
    • reduce/optimize operational cost
  • Collaboration between the NHS and the university is emphasized as a way to improve patient safety across many healthcare domains.
  • Without the Impact Acceleration Account funding, the project would not have been possible.

---

Detailed methodology / instruction-style elements (as described)

• Scale the discharge-letter audit process

  • Identify the safety-critical role of discharge letters (GP-only receipt, community-care coordination).
  • Quantify the operational challenge (150k–200k discharge letters/year; manual review takes 15–20 minutes each).
  • Target post-discharge adverse events, linking:
    • ~20% occur after discharge
    • ~60% linked to poor communication

• Build an interdisciplinary team

  • Assemble: clinicians + computer scientists + patients + quality improvement teams.
  • Hold initial alignment meetings to agree on:
    • shared goals
    • what “quality” means for discharge letters
  • Bring together data and expertise (clinical criteria + technical implementation).

• Develop and validate AI-based auditing

  • Use large language models to:
    • detect patterns in discharge-letter text/data
  • Implement a validation step where:
    • clinicians review/compare the AI audit outputs to ensure quality.

• Operational deployment target

  • Automate review so the system can:
    • audit discharge letters at full scale (for all hospitals over a year),
    • dramatically reducing time compared with manual auditing.

---

Speakers / sources featured (as stated)

• Northern Care Alliance NHS Foundation Trust (organization speaking/represented)
• University of Stirling (collaborating partner)
• Impact Acceleration Account (funding source)
• NIH R (NIH R applications mentioned as being submitted)
• NCA postdoc (role mentioned; specific name not provided)
• Speakers (individuals):
  • No specific individual names are provided in the subtitles for the speakers.