The future of artificial intelligence in health is placed under the spotlight here by The Alan Turing Institute, UK
Artificial intelligence (AI) will have a profound, positive and transformative impact on the health of European and worldwide populations. There is huge interest and worldwide investment in general AI, and health systems are one of the most important application areas. Europe is uniquely placed to play a leadership role in the race to build better, more accurate and safer health technology systems through the use of AI. European researchers are at the forefront in the development of safe, robust and effective AI delivering substantive patient and public benefit.
If properly managed, AI research will deliver both sustained health improvements as well as economic benefits to Europe. It will give our clinical and health researchers the time to focus on what they do best, enhancing their skills and maximising their expertise by augmenting their activities through AI systems that provide pertinent and timely information and reduce the burden from more mundane tasks.(1)
However, in order to reach this potential and for Europe to lead in this field, we need to address five key challenges. Policymakers should focus their attention on delivering these.
1. Data as infrastructure – “getting the learning base right”. AI algorithms are ravenous for data. It has been said that if AI is a rocket then data is the fuel. Data fuel is needed both in high volume and high quality. Poor quality data leads to systems that underperform.
Any AI system is only as good as the underlying learning base as, once trained, an AI algorithm is simply a mathematical compression of the data that it’s learned from. In addition to data quality, AI systems inherit any systematic or sampling bias present in their training data. For example, AI models using genetic information for disease risk prediction may be prone to bias or loss of performance unless trained on biobanks that are fully representative of population diversity of ethnicities. We must ensure that the less well-off are represented in the learning bases. This being especially true of genetics, where we often see an under-representation of non-European recent ancestries.
The learning base is the critical first component of any AI system and, as such, it is imperative that governments invest in European federated learning bases, connecting up health data at scale that can allow algorithms to train most effectively across all representatives of society. We need to ensure we have captured the full diversity of population data that will be encountered in the real world or, if not, have in place the necessary validation procedures to ensure that learned algorithms are robust in their conclusions (see below).
Within the UK, Health Data Research UK (HDR UK) are tackling this problem head-on. As Professor Morris, director of HDR UK notes:
“The UK has some of the most renowned data sources internationally, from NHS data to clinical trials, world-leading cohort studies and social, molecular and environmental data.
“Our ambition is to revolutionise research and innovation in the UK by working in partnership to enable linkage, access and analysis of these health data at scale and in a trustworthy way.”
Spearheading this is a series of Digital Innovation Hubs as exemplars of the technical and analytic challenges that lie ahead.
2. Public and patient engagement – “listen, learn and deliver benefits back to the people”. If AI is to transform European health services, we must bring the public with us along the journey. Public concerns regarding privacy and use of their health data for commercially driven activity are genuine and prescient. Important recent developments that need to be pursued include “algorithmic fairness” and “explainable AI”, the former ensuring that AI-based recommendations do not discriminate across key characteristics, such as race and gender and the latter building systems that provide interpretable recommendations.
3. Engineer robust solutions – “reproducible performance and validation of systems from bench to bedside”. Health data systems need to be robust and trustworthy. Due to the opaque nature of many AI algorithms, such as deep learning used for medical imaging diagnosis, the validation of such systems is more nuanced than testing traditional computer programs where formal system checks can be implemented. Care is needed in the design of AI validation protocols.
These protocols need to be explicit, auditable and in place at the study design phase and prior to the training of the algorithms. “Reproducible research”, whereby external partners can rederive claimed performance results is essential and will need to become a mantra of health systems AI. The Alan Turing Institute in partnership with UK and international partners have been developing guidelines in this respect(2).
In engineering solutions across Europe, AI systems are being developed in the leading hospitals by the leading researcher units using the best measurement technologies. We need to understand how performance is affected when moving from state-of-the-art research labs to real-world clinical settings, such that we design systems to be robust with well characterised and trusted performance measures.
4. Regulation – “delivering safe algorithms”. Ensuring that algorithms meet pre-defined acceptable levels of performance and safety guarantees, with smooth degradation of performance if applied to new environments requires new policy. Current regulatory procedures for software in medical systems are based on an assumption of static, explicit, code. Yet, by their very nature, machine learning systems are designed to improve their performance through continuous exposure to data. For continuous learning systems, we will need effective measures for monitoring and updating and clear, auditable requirements when systems go wrong.
5. Invest in the foundations – “why it works”. Modern AI is in its infancy, driven by an explosion of data captured at scale coupled to computers now able to process the vast number of numerical computations needed to train new classes of models. AI has shown impressive, game-changing, empirical performance but we still have little understanding of the theoretical foundations of why AI works so well and where it might be prone to brittleness. There is a danger of building our AI house on the sand. Fortunately, Europe is home to some of the leading computer scientists and computational statisticians working on the foundations of AI.
We have seen recent investment in national institutes, such as The Alan Turing Institute in the UK and the Finnish Centre for AI that are addressing fundamental research in AI. These centres are providing a focus for researchers developing the core principles and new algorithms for machine learning at scale(3). Europe must join together to combine our strengths.
The newly established European Laboratory for Learning and Intelligent Systems (ELLIS) is a prime example of how Europe can lead alongside our international partners. In the words of Bernhard Schölkopf, director at the Max Planck Institute for Intelligent Systems and co-founder of ELLIS: “Applications of AI are changing our lives and these changes originate from centres where the best research is done. Europe should play an active role in supporting fundamental research in machine learning in order to influence this transformational process. ELLIS will ensure that such research is performed in and shaped by the open societies of Europe to deliver the benefits of AI to all.”
If Europe invests in these five areas then the future is bright. New algorithms will have access to open, secure data learning bases, representative of European societies. The public will be informed and kept abreast of the benefits that accrue from donating health data at the population scale. Engineering principles and regulatory protocols will provide clear guidance to AI researchers and AI start-ups on the path needed to take ideas from concept to deployment in the clinic.
The leading AI research labs, connected and enhanced through structures such as ELLIS, will provide the foundations underpinning all such developments and bring on the next generation of young research leaders and disruptive AI technologies that will drive forward improved health outcomes. If we succeed in this endeavor then the potential to improve health for all Europeans will be dramatic.
Professor Holmes is Scientific Director for Health at The Alan Turing Institute, the UK’s national institute for AI and data science and chair in biostatistics at the University of Oxford. His work is supported by Health Data Research UK, the Li Ka Shing Foundation, the Medical Research Council UK and EPSRC UK.
1 Deep Medicine. (2019). Eric Topol
2 “Machine learning and AI research for Patient Benefit: 20 Critical Questions on Transparency, Replicability, Ethics and Effectiveness”, S. Vollmer et al (2018), https://arxiv.org/pdf/1812.10404.pdf
3 For example, Scalable Nonparametric Sampling from Multimodal Posteriors with the Posterior Bootstrap, Edwin Fong, Simon Lyddon, Chris Holmes. (2019). PMLR 97:1952-1962
Chris Holmes
Professor
University of Oxford
The Alan Turing Institute
Tel: +44 1865 285874
chris.holmes@stats.ox.ac.uk
www.stats.ox.ac.uk/~cholmes/
*Please note: This is a commercial profile
