Sunny Dosanjh, Director lead for Deloitte’s Healthcare AI and Data team, explores the three key aspects in scaling AI in healthcare
Healthcare Artificial Intelligence (AI) is a challenging field to enter, with a lack of widespread commercial success. This is not the case with other industries such as Financial Services which has seen ventures into AI go from strength to strength. Why, given the relative transferability of the technologies, have we not seen widespread successful implementations in Healthcare and the NHS? More importantly, how do we get it right going forward? We think there are three key steps:
1. Understand every inch of the regulations
The regulatory environment for Healthcare is complex and ever-evolving. Innovators can find themselves lost in the maze of Software as a Medical Device (SaMD), GDPR and confidentiality rules, but this need not be the case. These regulations are only intended to protect patients rather than stifle innovation. This means two simple steps can be taken to make sure innovation makes it to the frontlines.
- Invest in specialist regulatory/ compliance expertise: Too often we see cases where early-stage products look to regulatory expertise on the sudden realisation that they have entered into forbidden space. Unfortunately, by then, it is often too late. Instead, experts and the relevant systems/processes must form a key pillar of AI development. Be careful of the old innovation adage of failing fast, whilst a key pillar to development, failing on regulations is simply a never event.
- Don’t become a part of the regulatory landscape (until absolutely necessary): Innovators must make deliberate and carefully planned decisions to ensure they are only running into the constraints of the regulations if absolutely required. For example, in our collaboration with eConsult, where we have built AI triage for eConsultations to reduce administration burden[1], we used an in-depth understanding of the SaMD regulations to make the decision not to prioritise the clinical task list. Whilst this would clearly benefit administrative triage, it would also have tipped the innovation into a medical device, bringing with it the additional costs and timelines which may have prevented us from effectively entering this market.
2. Put impact, not technology, at the centre of your solution
We have all been there, focussed on the amazing accomplishment of our AI algorithm or our shiny user interface. Whilst this novel technology can seem impressive, it doesn’t mean it results in tangible benefits to clinicians, patients, or the NHS. It is these tangible benefits that will determine whether your product will scale or fail (even if the technology is super simple). The issue of putting technology ahead of impact is epitomised when looking at the 100 most referenced papers on the topic[2]. Only 11 per cent of these papers involve solutions where the benefits were well defined enough to move to a clinical trial phase. This lack of focus has led to very few (if any) real-world cases of AI being used at scale across the entire NHS. NHS-X/NIHR have been paving the way by funding 80 technologies through their flagship AI Award[3].
This challenge is further exacerbated with healthcare not always having well-defined success measures (e.g. in private businesses, you can measure profit, whereas in healthcare you need to consider financial, health, patient experience and much more). Therefore, it is pivotal to engage with on-the-ground expertise early, this proved critical in the adoption of Deloitte’s Referral and Intelligent Triaging Assistant (RITA) solution, itself an NHSx AI Award Stage 4 winner[4]. By thoroughly understanding how Consultant doctors were triaging GP referral letters and the impact this had on their time, we were able to design a product that not only aligned to their clinical safety approaches but also frees up clinical time for direct patient-facing activities by removing administrative burdens.
3. Show your working
Trust is integral to any implementation of AI within healthcare. With the consequences of mistakes being high, innovators cannot approach the market with an assumption that their model’s efficacy data will simply be all the clinicians and managers need to put the solution in to action. Instead, healthcare innovators have a responsibility to be proactive and transparent in demonstrating how their product works. Essentially, AI in healthcare must be explainable (to the masses).
- Build with explainability as the primary design principle: Most AI technologies focus on efficacy first, explainability second – we need to switch that importance around. This means when companies consider their MVP, they need to ensure it is viable through explainability. Yes, it may take longer to build, but a complete “blackbox” ML solution will not get the buy-in to scale.
- Use ground truths: To truly assess the effectiveness of AI, we need to be sure we are comparing it against the right benchmark. For example, most skin imaging AI solutions use a clinician’s viewpoint on whether a skin image should be referred to or not. This has several challenges: what if the clinician is not an expert, what if a different clinician would disagree and what if they are wrong? Ultimately, the only way to really test this AI would be with the skin pathology, giving the true position on whether the referral was required or not.
- Change with the times: All solutions also need to consider how they will adapt as clinical guidelines/practice moves forward and provide clinicians with an assurance of this process. For example, machine learning solutions require large historic datasets and may need to pause (or completely stop) should clinical practice change, as their models become based on outdated clinical practice.
Making the most of the opportunity
Ultimately, Healthcare holds a wealth of potential. Beyond financial and efficiency benefits, Artificial Intelligence has the potential to deliver a widespread positive impact on society if done right.
References
[1] Ai in Health and Care Award – funded projects 2021 (eHUB). (n.d.). Retrieved November 9, 2021, from https://www.nihr.ac.uk/documents/ai-in-health-and-careaward-%20funded-projects-2021/27866#eHub
[2] Sreedharan, S., et al. “The top 100 most cited articles in medical artificial intelligence: A bibliometric analysis.” Journal of Medical Artificial Intelligence 3 (2020): 1-12.
[3] Ai in Health and Care Award winners. (n.d.). Retrieved November 9, 2021, from https://www.nhsx.nhs.uk/ai-lab/ai-lab-programmes/ai-health-and-care-award/ai-health-and-care-award-winners/
[4] Ai in Health and Care Awards – funded projects 2020 (RITA). (n.d.). Retrieved November 9, 2021, from https://www.nihr.ac.uk/documents/ai-in-health-and-care-awards-funded-projects2020/25625#RITA:_Referral_Intelligence_and_Triage_Automation_-_Deloitte
Please note: This is a commercial profile
© 2019. This work is licensed under CC-BY-NC-ND.
