# Bespoke Large Language Models for Digital Triage Assistance in Mental Health Care

**Authors:** Niall Taylor, Andrey Kormilitzin, Isabelle Lorge, Alejo Nevado-Holgado, Dan W Joyce  
**arXiv:** [2403.19790](https://arxiv.org/abs/2403.19790)  
**Date:** 2024-03-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.19790v1/

## Key points

- There is a pressing need for efficient triage processes in mental health care to address long waiting lists for specialist services in the NHS.
- Large language models (LLMs) offer a promising solution for processing unstructured clinical data within electronic health records (EHRs) to improve triage decisions.
- A team of researchers from institutions like the University of Oxford and University of Liverpool have developed and evaluated three different approaches for LLM-based ingestion of variable-length clinical EHR data.
- The model is designed to provide triage recommendations aligned with existing clinical practices, be resource-efficient, and handle variable-length inputs at different levels of granularity.
- The study utilized electronic health record data from Oxford Health NHS Foundation Trust (OHFT) to train and evaluate the model's performance, aiming to contribute towards more effective digital triage assistance in mental health care within the NHS.

## Summary

In the field of mental health care, there is a pressing need for efficient triage processes to address the long waiting lists for specialist services in the United Kingdom's National Health Service (NHS). With between 370,000 and 470,000 new referrals into secondary mental healthcare services each month in 2023, clinicians face the challenge of quickly and accurately determining the most appropriate team to assess and treat patients based on information contained in electronic health records (EHRs). To improve this process, large language models (LLMs) offer a promising solution for processing unstructured clinical data within EHRs. By leveraging LLM technology, clinicians can efficiently ingest and analyze voluminous clinical notes to make informed triage decisions. This not only has the potential to reduce referral waiting times but also enhance the evidence available to support these decisions. In response to this need, a team of researchers from institutions including the University of Oxford and University of Liverpool have developed and evaluated three different approaches for LLM-based ingestion of variable-length clinical EHR data. Their model is designed to provide triage recommendations that align with existing clinical practices while being implemented on a single GPU, making it practical for use in resource-limited NHS environments. Key considerations in developing this model include its ability to handle variable-length inputs at different levels of granularity, such as documents, referrals, and individual instances. Additionally, the model is designed to be resource-efficient in terms of GPU compute and memory requirements, enabling repurposing of foundation LLMs for specific clinical use cases. Importantly, the model allows users to access evidence from source EHR data that informs triage recommendations in line with ethical guidelines on transparency in AI decision-making. The study utilized electronic health record data from Oxford Health NHS Foundation Trust (OHFT) to train and evaluate the model's performance. By addressing challenges related to data governance processes and ethical considerations surrounding LLM usage in healthcare settings, this research aims to contribute towards more effective digital triage assistance in mental health care within the NHS.