Author: Dr Paul O’ Reilly | Reading time: 8 Minutes
The resurgence of interest in Artificial Intelligence over the last 10 years has resulted in remarkable progress in many problem domains, such as autonomous vehicles, speech recognition, semantic search, and so on. In parallel, there has been a growing interest in deploying AI tools within the Healthcare system, to discover new treatments, assist clinicians in diagnosis, stratify patients, and improve outcomes for those patients, by improving the speed and accuracy of clinical decisions.
What is AI?
AI is a broad term for a range of technologies and approaches to simulate intelligent behaviour and reasoning in computer systems. Although in popular culture it is assumed to be the simulation of general human reasoning (or so-called Artificial General Intelligence), in most applications today, it is focused on usage with a narrower scope than AGI. Using AI in this narrow sense in some ways allows the development of useful tools in a broad range of domains while avoiding the more philosophical and unanswered questions around general reasoning.
AI is often used interchangeably with two other terms - Machine Learning (ML) and Deep Learning (DL). While there are links and overlaps between the three terms, it is important to distinguish between them.
Machine Learning is the term used to describe a broad array of learning algorithms whereby computers can ‘learn’ the characteristics inherent in their environment or data, give insights into the data and make predictions based on these data.
Deep Learning is a specific type of Machine Learning, using so-called ‘deep’ neural networks to identify patterns and features within the data, and is currently the major focus of AI research, since it has been shown time and again to be a powerful technique for leveraging data and knowledge into insight.
Opportunities for AI in Healthcare
There are many opportunities within healthcare for AI to assist scientists and healthcare professionals, in improving patient care and outcomes. Although clinical algorithms are undoubtedly important, and AI will play a large part in diagnosing and treating patients, their usage can be multifaceted.
Embedding AI seamlessly in intuitive and usable workflows can automate laborious and error-prone administrative tasks that are an inevitable part of a physician’s workload. On the other hand, an equally-important application is during the research and discovery stages for new biomarkers and treatments.
The ability of AI to deliver novel insights into large, heterogeneous and diverse datasets challenges existing statistical and bioinformatics approaches - and early use of AI in the discovery pipeline will also allow early adoption of these approaches when the treatments and biomarkers reach the clinic.
This end-to-end approach could allow precision medicine to be realisable at scale, and free up clinicians to concentrate on their patients rather than the intricacies of determining the best-individualised treatments for them, which often involves human decisions, and can vary from clinician to clinician and from patient to patient.
AI has the potential to have a major impact on Healthcare AI, particularly Imaging, Integration of Multimodal Data and Drug/Biomarker Discovery. In terms of maturity and adoption, Imaging appears to be a prime area in which to apply AI and there are a number of specialities in medicine which make use of AI for imaging analysis, including Radiology, Cardiology, Histopathology, and Cytopathology.
In Radiology, and also in the arena of image-assisted diagnosis of heart disease, the use of digital images, and image processing is accepted. This is evident in the maturity of the use of AI within the domain.
There are multiple platforms and applications approved for clinical use, and AI is becoming embedded in the workflows of radiologists. However, it appears that this AI is being used to augment rather than replace radiologists and we are still some way from the vision of Geoff Hinton, who stated in 2016 that "It's quite obvious that we should stop training radiologists now".
In the pathology domain, AI has been applied in a number of ways. In addition to the prime use case of assisting pathologist diagnosis and grading of cancers, a number of other applications of AI have been addressed, including screening of negative results (to pick up false negatives), characterization of tissue, and quantification of tumour.
While the digitisation of pathology is an ongoing process, the recent regulatory approvals for manual primary diagnosis using digital images by Philips and Leica, and subsequent approvals for AI-assisted pathology of prostate cancer and breast cancer (which requires digitisation) will undoubtedly accelerate the use of AI in this complex domain.
Outside of the direct, primary clinical use cases for imaging analysis in pathology, there are many more pathology tasks in the research and discovery domain, which would benefit from automation through AI. Indeed, reviewing the research literature, there are a multitude of applications of AI in the analysis of H&E, IHC and IF images. Often, AI is used to reproducibly automate tasks which pathologists and researchers find challenging and laborious to carry out manually.
Outside of imaging, there is also much development in the application of AI in ‘-omic’ data. Indeed, if ‘traditional’ machine learning approaches can be termed AI (which I would argue they can), then through the use of clustering, nonlinear regression, and classification approaches such as random forests or XGBoost, AI is probably more embedded in these areas than anywhere else in healthcare.
Given the amount and dimensionality of the data provided by techniques such as gene sequencing, RNA expression, mass spectrometry etc., analysis of such data would not have been possible without using AI to give true insight into the data. However, now we are seeing the use of more broadly-defined AI systems based on deep learning in these domains to provide deeper insights into correlation and causation, and this has the potential to further interrogate new and legacy datasets to obtain new insights, both in research and clinically.
In the drug discovery domain, as in others, the application of AI is becoming, if not common, more prevalent. This is not just using the imaging and genomics AI approaches already discussed, but there are now tremendous advances in applying AI to problems such as protein folding. Indeed the recent announcements from Google’s DeepMind regarding the groundbreaking results from their DeepFold algorithm, definitely suggest that AI will transform all aspects of drug and biomarker development.
Challenges for AI in Healthcare
Despite the many potential advances made possible by AI, there are a number of challenges to its adoption and use, some of which are shared with other domains, some of which are unique to the healthcare domain, and some of which overlap incompletely.
In building AI systems, the quality and quantity of the data used is massively important. This big data, and any associated annotations and/or metadata must be captured with its provenance tracked and available for audit. Given the sizes of the source data files (images, ‘omics’ data) and the number typically required for developing algorithms, it is essential to have scalable, efficient and reliable data management capabilities, for the development, validation and post-deployment phases. And, of course, such data management must be private, secure and auditable.
In the clinic, one of the obvious challenges is that of regulation. Although a number of AI solutions have started to be approved for use in day-to-day clinical practice, the regulators have, rightly, been concerned about a number of aspects of AI. These include reproducibility and reliability of results, when the hardware and architectures of AI solutions are ever-changing and learning is often a stochastic process, susceptible to the vagaries of stochastic randomness.
Given the large amounts of data typically used in developing AI, there is a need to trace the provenance of data used, and archive the parameters used. Much of the power of AI comes from the ability to update the algorithm in response to new information and data (so-called online training) but this is fraught with regulatory risk.
In 2018, the FDA published a discussion paper on AI algorithms as medical devices, and this post-deployment real time update of algorithms was discussed, but there does not appear to be a consensus as to how this should be handled, and current AI systems approved for clinical use are all ‘fixed’ algorithms.
One challenge to the adoption of AI is in the realm of explainability. Many current AI systems are essentially treated as a ‘black box’ which leads naturally to resistance to its use. If doctors and their patients do not know the ‘rationale’ behind the decision or advice from an AI, it naturally will lead to resistance to its use.
Ongoing work on examining the explainability of deep learning networks is helping address these concerns in what is the predominant ‘black box’ AI technology, and hopefully, the fundamental work being done on this will be translated into the AI systems used by doctors, and allow them to develop confidence in both the technologies and the results.
Finally, a major concern about AI systems, particularly as they are being more widely deployed is that of bias. At a minimum, it is necessary to ensure that the data used to train and validate AI algorithms includes no obvious sources of bias.
In order to do this, the data used to train the models, and any associated metadata must be stored and tracked reliably and securely. There is also the need to collect specific data sets in order to ‘challenge’ the AI and detect bias.
This data integrity problem is, of course, not sufficient as there are many non-data sources of bias (such as organisational and societal biases) which we also need to be aware of, but it is necessary if we are to have AI which does not bias towards or against specific groups. This is an ongoing concern, and any post-deployment surveillance needs to examine the performance of the AI in the real world.
This is a necessarily limited subset of the areas within which AI is being applied to healthcare. Whatever the use case, from natural language processing for the huge amount of unstructured text held in Electronic Health Records and the corpus of scientific literature, to robot-assisted surgery, the healthcare industry is finding ever more opportunities to leverage AI-powered technologies to assist and augment all stages of the healthcare spectrum.
The opportunities for AI in healthcare are huge, as discussed here. If researchers and industry can address the challenges also discussed, AI has a large part to play in improving the diagnosis and treatment of disease, and as a consequence has the potential to help improve the patient journey and outcomes.
Regulatory Challenges and AI