The “inconvenient truth” about AI in healthcare

The “inconvenient truth” about AI in healthcare

Original Paper

For the full text of the paper please click here

Blog Post - Behind the Paper

François-Marie Arouet, 18th century French author and iconoclast, better known as Voltaire, quipped, “The art of medicine consists of amusing the patient while nature cures the disease.” 

Though medicine has progressed in the intervening centuries it remains an art informed by science: both the art of bearing witness, of helping people find meaning in the maelstrom of life's immediate and existential challenges and the less appreciated art of managing uncertainty. It is in this latter regard that AI, broadly speaking, holds promise. Precision medicine, optimised systems and proactive population health have all been forecast but arguably more important is the potential liberation of the time and ingenuity of clinicians to do what they are uniquely able to do - to care for other people. Humans doing what humans do best. In reality this vision of AI in medicine is some way off and for now AI has been developed in focused areas where there is both relative consistency in the real world and a glut of examples present in the digital one. The dream for some and the nightmare for others is to reach escape velocity and transcend human levels of general clinical reasoning. There are plenty of proximate challenges to overcome however and many of the technical challenges in algorithm development have been well described. 

We have realised in our work in developing algorithms, in training health data scientists, in working with payors and providers of care in the US and in working with governments internationally that there is still much work to be done either side of the algorithm. By this we mean that, though it is known that AI requires data, there are vanishingly few accessible sources of curated healthcare data that reflect the broad variety of practice and are kept up to date. Furthermore the data infrastructure, defined as the hardware and software to securely aggregate, store, process and transmit healthcare data across a health system in real time and at scale is largely absent. Secondly, once algorithms are developed to actually deploy them in practice requires significant clinical involvement, change management expertise and long term organisational support. These factors are typically not discussed in research papers nor do they feature in conference presentations, though arguably collectively they determine the likelihood of research findings being successfully translated into clinical impact.  

It is our experience that clinician involvement at all stages of the algorithm development process is beneficial and will certainly make subsequent implementation easier, however we wrote this paper to kick-start a discussion around the following question: is data infrastructure a public health intervention?  We believe that it can reasonably be argued that it is and, with the benefits of hindsight, will be considered by future generations as one. Like all public health interventions it has the potential to create enduring benefit but will require not just a broad coalition of support and partnership between the public and private sector but also the trust and enduring support of patients.