Artificial Intelligence and supporting diagnosis in 21st Century Primary Care
Chaired by Brendan Delaney
Aims: Diagnosis is the primary task of the General Practitioner and diagnostic error the largest risk for both patients and practitioners. Interest has grown significantly in the role of Artificial Intelligence (AI) in solving complex problems. Is this the solution for GPs? Symposium Content AI describes the use of computational systems to replace all, or some aspects of human judgment. In Medicine, personal level decision support for patient management, is starting to appear, particularly the use of ‘Deep Learning’ models in image analysis and algorithms based on large scale clinical data for predicting clinical risks. The ‘Learning Health System’ (LHS) describes an informatics approach to managing the data and knowledge learning cycle as part of a distributed system. In the context of diagnosis, data are fragmented in place and time and both incomplete and potentially biased. Learning is therefore inefficient, and tends to learn the ‘biases’ and models cannot be applied reliably to new data. We will describe an approach bringing together tools to improve the structure and capture of coded data during the consultation, Machine Learning and decision support to create an ‘LHS for diagnosis’. Format of symposium Three 15m presentations, and a 15min discussion, will describe the data capture requirements, the knowledge management requirements and the potential impact of decision support on clinical decision making. Rationale for bringing these presenters/presentations together. Speakers consist of a GP with a Chair in Informatics, a Psychologist with a focus on clinical judgment and a Computer Scientist with expertise in clinical knowledge management. The joint programme of work has recently received further funding from Cancer Research UK and the Imperial NIHR Patient Safety Translational Research Centre. Intended audience GPs and clinical epidemiologists with an interest in diagnosis. Plans for discussion / interactive element Potential topics for discussion include: What can be done to improve the transparency of AI? Will AI replace the GP? Will AI lead to sharing of diagnostic knowledge or a growth in proprietary systems? What is the role of the patient and the public? How can we measure improved diagnosis?
Presentation 1 - Artificial Intelligence in diagnosis, more than just an algorithm.
Professor Brendan Delaney Imperial College London
The implication of a ‘Learning System’ is that knowledge alone cannot improve diagnosis, a clinician must take action on the basis of that knowledge, creating new data to further develop the AI, thereby closing the learning loop. Significant challenges in data standards, maintenance of clinical meaning, data linkage, data preservation and privacy need to be overcome at scale. If a system is to support ‘generalist’ care, it needs to support all differential diagnoses, not just certain conditions such as diabetes or cancer. Access to data and the ownership of the knowledge created are going to be critical issues for the future.
Presentation 2 - Clinical decision aids - friend or foe?
Dr Olga Kostopoulou Imperial College London
Overconfidence in our intuitive judgements makes us less likely to consult helpful decision aids, despite evidence that using an algorithm or clinical prediction rule performs better than unaided judgement. Doctors are reluctant to use diagnostic decision aids, reserving them for the ‘most difficult’ cases or the inexperienced trainees. But are doctors able to recognise when a case is difficult and when they need help, in other words, are they accurately confident? The talk will discuss some of the issues and concerns surrounding the use of clinical decision aids, and will present some evidence of effective ways to support clinical diagnostic judgements.
Presentation 3 - Beyond big data – knowledge management for the Learning Health System
Dr Derek Corrigan Royal College of Surgeons Ireland, Dublin.
Achieving the goal of the LHS to facilitate the iterative application of clinical knowledge requires a consistent knowledge management platform capable of representation, curation, and retrieval of clinical knowledge in a federated digital library. The tracking of how knowledge is created, versioned and applied in practice is an essential requirement to ensure trust and address ethical, security and data protection concerns regarding use of AI more generally. The talk will discuss what such a knowledge platform might look like and possible solutions to ensure trust.