Large language models (LLMs) in adult social care

Abuse in social care can cause severe harm and is challenging for care providers to detect due to the vast amount of case record data. This project explores the feasibility of developing a large language model (LLM) to help social workers identify risk factors for abuse within local authority case records.

Social workers manage large caseloads, often containing over a million words of text, making it difficult to spot crucial information. This six-month study will assess available data, secure access, and prepare for future development of the model. The proposed tool would highlight key details—such as family relationships, financial issues, or past allegations—rather than predict abuse, ensuring compliance with General Data Protection Regulation (GDPR).

By integrating into local authority systems, the model could support better decision-making and help prevent abuse by drawing attention to critical case information that may otherwise be overlooked.
 

• Aims

  The primary research question is: How feasible is it to develop and implement a large language model that can highlight information relevant for abuse in local authority administrative records?

  The project aims to:

  • Identify data sources and establish access to records necessary for model development.
  • Engage stakeholders to understand the structure and availability of relevant data within
    local authority systems, and explore ethical issues, public perception and potential
    challenges around bias and trust in AI.
  • Plan the technical and ethical framework for the implementation of the predictive
    model.
  • Conduct a narrative literature review to assess existing prediction models and inform the
    study.

   

• Methods

  The project will be divided into four key activities:

  1. Workshops for stakeholder engagement

  We will conduct workshops with local authority partners and other stakeholders to  identify where abuse-related information is recorded in case notes and establish the availability of relevant data, such as demographic information, structured needs assessments, and free-text records. Discussions will focus on understanding how data is recorded across systems and how these change over time. We will also explore the feasibility of focusing on specific types of abuse, such as financial or physical abuse, and whether certain demographic groups, such as minority ethnic populations, are sufficiently represented for meaningful analysis.
  Additionally, we will engage with local authorities and continue to engage with public involvement groups to discuss ethical issues, public perception, and trust in AI. This engagement will help us understand community concerns, reassure individuals that open-source AI models can be run locally so their data need not be shared, and ensure that our approach aligns with ethical standards and public expectations.

  2. Securing data access

  The goal of the workshops will also be to secure data-sharing agreements with one or more local authorities, so that the necessary data is available for developing a model.  We will also address in these documents potential legal, ethical, and intellectual  property concerns, securing a memorandum of understanding and data access protocols. We will be guided by the outcome of the workshops in our application to the appropriate ethics committee regarding the creation of an AI model.

  3. Implementation planning

  A detailed implementation plan will be developed, outlining the steps required for integrating the predictive model into local authority systems. We will explore whether integration into existing IT systems is feasible or if a stand-alone solution is necessary. This plan will include technical, practical, and ethical considerations, along with potential challenges and mitigation strategies.

  4. Narrative literature review

  We will conduct a narrative literature review to synthesise existing research on predictive models in social care and the application of technology to safeguarding issues.  The review will inform the development of the model and help ensure that the approach builds on prior research. We will also connect with similar studies, such as the SSCR project at Leeds Beckett, to explore opportunities for collaboration and knowledge exchange.
   

• Planned outputs

  Proposed outputs:

  1. Secured data-sharing agreements and memoranda of understanding with local
  authorities.

  2. A narrative literature review that summarises the state of current knowledge and
  positions the proposed model within this context.

  3. A detailed implementation plan outlining how the model could be developed, integrated
  into local systems, and evaluated for effectiveness and ethical compliance.

  4. Ethical approval from the relevant ethics committee.