AutoConfidence: Automatic confidence assessment for deep learning in healthcare

At the beginning of my fellowship...

Clinical implementation of artificial intelligence (AI) methods is a challenging problem, particularly in high stakes environments such as radiotherapy, where the risk to patient safety, in case of error, is high. As many CE marked medical devices now contain AI technology and operate as ‘black boxes’ from the clinician’s point of view, my primary objective was:

to independently estimate per-patient AI confidence, to allow patient and clinician to have appropriate clinical confidence in every AI assisted decision.

This required development of a novel AI confidence estimation method, dubbed ‘AutoConfidence’, alongside establishing a framework for building confidence in AI methods through robust evaluation, validation and implementation using an ISO9001 based quality systems approach.

During my fellowship...

Through the Topol Digital Health Fellowship, and the thanks to the support of HEE and the NSHCS, I have been able to develop AutoConfidence as a method for independently estimating the confidence level of an AI prediction, on a localised basis, allowing clinicians to identify areas of concern in image-based predictions and make safe clinical judgments in radiotherapy.

Importantly, this method works on a per-patient basis, without ground-truth information, so can be applied to every clinical case in routine practice, improving clinical confidence and hence safety and efficacy.

Through HEE, I have been involved in working with colleagues at NHSX to develop ideas for clinical confidence in AI into a general approach for AI in higher stakes environments.

Subsequently, I have demonstrated AutoConfidence for supervised and unsupervised learning in medical imaging applications, for both classification and regression problems, demonstrating its potential applicability to any AI-based method for image procession.

AutoConfidence and the surrounding deep-learning governance framework (dQMS), which has evolved from the original project, have been presented at national and international radiotherapy and radiology meetings, including ESTRO 2020 and the IPEM AI in RT and AI in MRI meetings. There are two publications in press and preparation resulting from this work. We are working with a commercial provider to incorporate AutoConfidence into their A- enabled product, for independent per-patient validation.

Personally, my career development has benefited greatly from this fellowship, which has enabled me to make contacts and progress this work rapidly, leading to a new post focused on AI research and clinical implementation.

Lessons learned:

  1. Protected time to explore a cutting edge proof-of-principle idea has enabled this project. Without this fellowship, the time would simply not have been available.
  2. Doing a technical project within the context of a fellowship program about digital technology implementation has been fascinating and shaped my approach towards making an implementable solution.
  3. Working with colleagues in a diverse range of areas and ‘digital’ experts from outside the NHS is a powerful combination.
  4. Considering the end-user and use case at the first conception of an idea is critical to translating out of the academic bubble into clinical and commercial environments.
  5. People before tech!