Establishing which modalities of artificial intelligence (AI) for early detection and diagnosis of cancer are ready for implementation in primary care: a Scoping Review

Talk Code: 
C.9
Presenter: 
Owain Jones
Co-authors: 
Owain T Jones, Kethaki Prathivadi, Smiji Saji, Charindu KI Ranmuthu, Natalia Calanzani, Jon Emery, Willie Hamilton, Hardeep Singh, Niek de Wit, Stephen Duffy, Fiona M Walter
Author institutions: 
University of Cambridge, University of Melbourne, University of Exeter, Baylor College of Medicine, Queen Mary University London

Problem

Approximately 360,000 people in the United Kingdom are diagnosed with cancer each year and approximately 160,000 people die of the disease. Most people diagnosed with cancer in the UK first present in primary care, where General Practitioners will evaluate (often vague) presenting symptoms and decide on an appropriate management strategy. More accurate triage of these presenting symptoms could lead to earlier diagnosis of cancer, and improved outcomes for patients, including improved survival rates. There has recently been huge interest in the application of AI technologies, including machine learning, to medical diagnosis. However, there are currently no AI technologies that are established in routine clinical care.

Approach

We performed a scoping review, aiming to explore and map the research landscape for AI technologies designed to aid the early detection of cancer, focusing on technologies which would be suitable for implementation in primary care settings. We searched Medline, EMBASE, SCOPUS and Web of Science bibliographic databases from the 1st January 2000 to 11th June 2019 for relevant published studies. We identified 10,456 relevant studies after removing duplicates, and subsequently assessed 793 full text articles. This led to 250 studies included in the quantitative synthesis.

Findings

The AI technologies we identified fell into three main categories: (1) applied to electronic health records and routine blood results, (2) applied to superficial digital imaging, such as lesions on the skin, buccal cavity or cervix, and (3) applied to subsurface digital images, such as ultrasound and thermography. We found a diverse range of AI techniques used in the studies, although neural networks and support vector machines were by far the most common. We also identified a range of outcomes measures used across the studies. Most studies represented very early stage translational research, using various AI techniques to address clinical problems using freely available online datasets. However, there appeared to be a paucity of clinical input into the design of these AI-technologies, and they were rarely validated in independent datasets or prospectively tested in clinical settings.

Consequences

Many of the AI-technologies we identified could potentially aid the early detection of cancer and other serious conditions in primary care settings. However, the research is currently at an early stage, and there are a number of further steps needed before the technologies can be safely and effectively implemented into routine primary care. One key step will be validation of the AI-technologies in independent datasets and prospective clinical studies set among the intended population. Health economic assessments, and research involving patients and clinicians are further important steps, to identify their opinions on AI technologies and potential implementation barriers.

Submitted by: 
Owain Jones
Funding acknowledgement: 
This research has been supported by an NIHR School for Primary Care Research Seed Corn Award for author OTJ. Additional funding has come from the NIHR Cancer Policy Research Unit, and the CanTest programme funded through a CRUK catalyst award