Artificial Intelligence to support cancer early diagnosis in general practice. (AI-DIP)

Research question This proposal aims to validate and evaluate in the NHS Artificial Intelligence (AI) to enable the early diagnosis of cancer in general practice.

Background Misdiagnosis in general practice, defined as a failure to enable definitive diagnosis of a patient within a prompt timescale, has significant repercussions for the care and survival of patients presenting with possible cancer.

Differential diagnosis lists presented at the start of a consultation can improve diagnostic accuracy, but this is best deployed before the GP consultation.

Recent developments in the USA with SOAP Health s Ideal Medical AI Assistant (IMAA) with the 'Perfect Medical Interviewer' (PMI) that can enable symptoms to be gathered pre-GP access (by an animated interactive virtual human) and seamlessly integrated with other data about that patient to provide a differential diagnosis list for the GP (RiskVue module).

In addition, cutting edge developments in Machine Learning (ML) from multi-modal data including natural language processing (NLP) and Answer Set Programming will be incorporated over time to strengthen the knowledge engine of RiskVue. SOAP Health is currently TR5 in the USA and we will take this to TR8 in the UK NHS.

We will use pancreatic and lung cancers as exemplars, as both carry a very poor survival rate on account of the late stage in diagnosis in the majority of patients. Both act as good exemplars for learning from multimodal data. Aims and objectives. Objectives 1.

To validate in the NHS a virtual-human medical history assistant developed by SOAP Health in the US to collect structured medical history data from the patient prior to the GP consultation, with an emphasis on usability across disadvantaged groups. 2.

To explore adaptation of the diagnostic risk model to the UK NHS by multimodal data analytics based on natural language processing and knowledge graphs (ILASP) to risk-stratify patients for early cancer presentation. 3.

To fine-tune the risk model to UK patients by exploring methods for representing changes in symptom and clinical data over time. 4. To integrate the learning system within a demonstrator NHS Integrated Care Board 5.

To evaluate the within-consultation medical assistant for supporting GP decision-making with reference to its impact on reducing inequalities in time to diagnosis. 6. To support entry to market of SOAP Health in the NHS. Methods Work will be distributed over five WS. Consisting of: 1. Adaptation and validation of the IMAA in the UK 2.

Multimodal ML to generate explainable and extensible UK-specific models for cancer risk. 3. Clinical evaluation of the revised UK Riskvue module 4. Co-working with NHS Sussex in developing the system in an NHS England Digital Innovation Lab. 5. Patient co-design, management and IP protection for NHS deployment.

Timelines for delivery PMI 12 months, RiskVue 36 months Anticipated impact and dissemination This proposal has the potential to transform the process by which GPs are supported in making accurate and timely diagnoses, initially for cancer, but also the many other conditions that are part of the wide differential diagnoses for lung and pancreatic cancer presentations.