Developing safe conversational artificial intelligence for specialist NHS neurology services

Aims: to develop safe clinical artificial intelligence (AI) assistants and establish generalisable approaches to their production and validation. I will develop an assistant to perform triage and decision support for people on the waiting list for neurology outpatient review due to persistent symptoms following head injury. Background: newly developed large language models (LLMs) perform well on medical benchmarks, and show promise in taking clinical histories, reasoning and suggesting management. However, consumer LLMs based assistants are not validated for use clinically. This is because the technology is new, there is no agreement yet about how their clinical validation should be approached, nor has validation been done for specific use cases. What's needed is targeted development of clinical AI assistants, led by experts clinicians and people with lived experience, working with regulators and policymakers. Enhanced triage and decision support after head injury is the ideal initial problem to address, solving an important issue and answering general questions about how we develop/validate clinical AI assistants. Clinical information is highly accessible via the history, head injury is common (~1 million UK ED attendances annually; 20-50% with persistent issues), there is a paucity of head injury neurologists (6 months) and prompt treatment is important. Methods and timelines for delivery: in-silico development of the clinical AI assistant for head injury triage/decision support development will begin immediately, including PPIE-informed interface design (WP1, Helix Centre, Imperial) and technical optimisation in simulated consultations (WP2). I will develop the architecture with Prof Barnaghi (a clinical machine learning expert, Imperial) and optimise for quality, safety, tone, privacy and non-bias. Voice-based interactions will be explored with Pearse Keane (Professor of Artificial Medical Intelligence, UCL). Next, a mixed-methods feasibility study (WP3) will be conducted in patients awaiting specialist head injury neurologist review at Imperial and NHNN (n=100). Performance will be evaluated by domain experts quantitatively, and via semi-structured interviews with users. Regulatory approvals (Hardian Health) and policy issues (Prof Niels Peek, Cambridge) will be considered (WP4), informing an effectiveness study following the fellowship. Generalisable methods will be shared via the Turing Institute, including new UK guidelines (WP5). PPIE: this proposal was informed by people with lived experience of head injury. I will form a PPIE group which will meet regularly to guide the design, development, evaluation and dissemination of outputs. Training and Environment: I will develop my computational skills with Profs Barnaghi and Keane (my mentor), train in qualitative methods (Oxford; Imperial) and clinical trials methodology (Imperial CTU). I will benefit from a funded PhD student (Imperial) and major existing investment in clinical AI (Imperial-X); with support from Microsoft ($250K), placing me optimally to become an independent leader in clinical AI. Impact: I will establish a general methodology to develop and evaluate clinical AI assistants, while taking a key step in applying this to help people after head injury receive prompt and appropriate treatment. This will answer pressing questions about how to maximise the benefits of clinical AI in the NHS, with transformative potential.