Realising the Potential of AI in Healthcare
The author (Dr Sarah Qian) in a specialist doctor in endocrinology and is a current MBA candidate and Service and Society scholar at London Business School. She believes in digital solutions and the power of AI to build resilient and sustainable healthcare systems. This article is the first of a two-part series examining the opportunities and challenges of investing in Healthcare AI.
Photo by Moritz Kindler on Unsplash
Globally, healthcare is in crisis. COVID was the spark that set the house on fire, but in many ways the damage is as much related to underlying structural problems as to the pandemic itself.
Without innovation to help us work smarter, the outlook is bleak: an already depleted workforce with record burn out, delays in chronic disease management, a precarious mental health burden.
Faced with this challenge, the one saving grace has been the progress in HealthTech, and in particular, AI in healthcare. Through necessity, the pandemic has pulled all stakeholders - patient, providers, industry - onto the same page. But where are we at exactly? In the first part of this series, I examine the current state of AI in healthcare and discuss strategies to bridge the gap between industry and providers to accelerate the translation to clinically meaningful impact.
What is AI?
In the current climate of rapid HealthTech innovation, it can seem like every app and device is promising AI. But what is the definition exactly?
AI is a broad term for the creation of intelligence mimicking human cognitions (like learning and problem solving) that respond autonomously to changing environments. Subfields exist based on how this learning is achieved (machine learning, deep learning) or by function (natural language processing).
Machine learning uses algorithms to recognise patterns. This could take the form of supervised, unsupervised or reinforcement learning, but all require the problem to be clearly defined. This technique has been advanced by the data revolution and forms the basis of many early AI ventures.
Deep learning goes further, taking advantage of increased computational power to layer neural networks, enabling systems to learn on their own, using only raw features as input. A high-profile example of this is DeepMind’s AlphaFold, which uses this technology to predict protein structures. For a fascinating (and accessible!) walk through, see Prof Graepel's lecture here.
Of note, these definitions exclude simple digitisation and automation from pattern recognition. While these can be beneficial to healthcare processes and enable AI delivery, they don’t generate insights on their own and so shouldn't be classified as AI.
Where are We At?
The events of the last two years have seen increased investment and growth in the AI healthcare industry. In 2021 the global AI in healthcare market size was estimated to be $6.9B, and is expected to grow to $67.4B by 2027 (CAGR 46.2%).
This energy, enabled by technological advances and coupled with rising demand, makes for an intriguing mix. I am most excited about the very real applications we are now seeing right across the healthcare value chain, all of which increase workflow efficiency and relieve pressure on existing services. Broadly speaking, these can be categorised into six categories - from initial consultation, to management in the community (see diagram). Next, I dig a little deeper into three of the most promising and dynamic areas.
Watch this Space
AI in Triage and Diagnosis
Symptom checkers are evolving from using simple pattern recognition to more complex algorithms capable of making diagnoses. At a basic level, eConsult’s eTriage is being used in NHS A+E departments to flag sick patients and suggest next steps. Similarly, K Health in the US provides basic diagnoses and lets patients know when care needs to be escalated.
As a benchmark of how far the technology has come, we can look at the research done by Babylon, a UK born digital healthcare unicorn. The Babylon Triage and Diagnostic System is based on a Bayesian network and when tested on a set of medical scenarios of patients, the system achieved similar sensitivity and positive predictive value as 7 real world doctors. There are still many unknowns - how will the system fare against real world patients with multiple problems? What would a larger sample size and safety data show? Nevertheless, this is a step up in accuracy compared to previous studies.
These applications provide a glimpse of what can be achieved, but we are still far from being able to replicate a clinician in our smart devices. There remain key differences in the way that clinicians work and the way machines learn that have yet to be overcome – for example, evaluating the subjective components of a patient history and weighing this alongside non-diagnostic results. However, advances in natural language processing (NLP) and quantum cryptography will allow medical data to be tapped and categorised more efficiently, so expect to see dramatic improvements in this space.
AI in Diagnostics
AI is a natural fit for the systematic analysis required to detect imaging abnormalities and this is a growing area backed by evidence. A study across the US, UK and Korea demonstrated that a deep-learning algorithm performed better than radiologists in the detection of breast cancer in mammograms. We are now seeing a wave of start-ups addressing this need, from X-rays (Radiobotics, Gleamer) to whole body imaging (Quibim).
The scope of this technology is increasing and we are seeing this being used in other imaging-based diagnosis, most notably in skin cancer. When shown photos of skin lesions from smart phones, Skin Analytics ‘DERM’ tool was shown to be comparable to specialists in the diagnosis of melanoma.
With clear proof of concept, we will see increased clinical uptake of these technologies with the challenge being successful integration into work flow. As detection by algorithms risk over-diagnosis, they would ideally be deployed alongside radiologists, relieving workload volume and allowing clinicians to focus on more complex cases.
AI in Therapeutics
The pipeline for AI driven drug discovery is growing. As in diagnostics, companies are leveraging NLP -driven insights from existing literature, biological data, and commercial information to more efficiently find drug targets and treatments. Some have chosen to focus on specific areas (HealX and rare diseases) and many are partnering with pharma to collaborate on clinical trials and commercialisation (Benevolent AI and AstraZeneca, Iktos and Pfizer). In the early start-up space, BaseImmune applies AI differently, predicting mutations in viruses to give a head-start on targeted vaccine production.
A crucial piece of the puzzle is being able to accurately manufacture drugs to specification. And here I return to DeepMind's AlphaFold, which enables the prediction of 3D protein structures from amino acid sequences to inform targeted drug design. Incredibly, DeepMind has released the source code and database of known proteins free of charge, accelerating research efforts around the world. With such dramatic progress, we are closer than ever to realising personalised medicine.
As we have seen, AI has much to offer the healthcare landscape. However, many challenges remain. And even before we get into the details of data engineering, and workflow integration, it’s important to take a step back and consider how to build trust and acceptance of the technology in order to facilitate partnerships between stakeholders.
Traditionally, medical professionals have been, at best, mixed in their acceptance of new technologies. The pandemic has provided a strong impetus for change, but a fatigued workforce remains wary.
I believe an ideological shift is needed. Perhaps it's time to ask ourselves what the alternative is, what should be the comparison?
As a doctor, I have learned to set a high threshold of proof – Primum Non Nocere, first do no harm. As such, the usual the point of comparison is best evidence. But in our current resource-limited setting, I challenge my colleagues to consider the alternative to be no access, a current reality for many. With this public health perspective, we widen the definition of harm to include those denied care. And given this, do we really have a choice whether to embrace technology that can help our systems work more efficiently?
One of the key concerns about AI in healthcare is the lack of human oversight, especially when the consequences of being wrong are so high. To address this directly, Explainable and Causal AI is a growing field providing visibility on abstracting features in the layers of neural networks in algorithms. This allows doctors to analyse how the AI reached conclusions from inputs and make adjustments as needed.
As a further layer of protection, solutions are being built for “out of distribution detection”, enabling the AI system itself to recognise outliers and refer them to humans rather than taking a guess on its own. Developments like these are vital mechanisms to enable AI and clinicians to work confidently together.
Finally, the industry needs ethical and regulatory frameworks that are both clear and adaptable to rapid innovation. This concerns the management of patient data, but also the classification of Software as Medical Devices (SaMDs) to ensure regulation and safety standards are met. It's great to see the Ada Lovelace Institute's initiative outlining Algorithmic Impact Assessment to measure the impact of these technologies and to guide governance.
With such dynamic growth and breadth of opportunity, there’s a lot to be excited about in AI and healthcare. To realise this potential, I think it's helpful to conceptualise a framework with short, medium and long term goals, like the one developed here.
Firstly, in order to further bridge the gap between industry enthusiasm and provider caution, partnerships should be formed early to build functionality with interoperability in mind.
In addition, there needs to be a focus on demonstrating reliability by using high quality data and developing safety evidence. This will increase user confidence and allow the technology to scale and integrate more broadly.
As for Investing...
Beyond the issues already discussed, AI in Healthcare is a challenging field for investments given the risk profile. In the next article in this series, I will expand on these risks and discuss specific considerations for early-stage investments in a speculative market. Stay tuned!