“AI” we hear it everywhere, with frenzied excitement. But what is it, and does it warrant the hype? Often overlooked, AI is not yet an “it,” despite almost universally being talked about in the singular. AI (artificial intelligence) is a group of diverse disciplines, which collectively may (one day) come together to present human-like intelligence capabilities in machine form. AI comprises statistics, machine learning, robotics, computer vision, natural language processing, chip design and many other computer science and mathematical disciplines, all fuelled by massive (and increasing) quantities of digital data and on-demand cloud computing.
The prospect of a singular, general AI remains a future ambition. Today applications of AI technologies are generally narrow, that is, they address to focus on a specific problem using a selection of techniques from the AI range. Of these, machine learning is particularly widespread. Machine learning (“ML”) is the practice of combining statistical methods and computational resources to create a mathematical model usually learned from existing data. The algorithm (computer code recipe) that generates the model can continue to refine the model as new data is added without needing additional programming, hence the concept of machine ‘learning’. Having such a model means you can forecast, theoretically with ever-increasing accuracy, a task that can be a least complex and time-consuming for humans.
“We can envision a future world in which condition-specific AI models trained on the world’s collective data are widely available”
In the context of providing cost-effective and scalable healthcare services, machine learning has huge potential. Academic and commercial researchers worldwide have reacted over the last decade or so with a myriad paper demonstrating detailed healthcare applications. Examples include: Google Deep Mind and Moor fields Eye Hospital, UK, used deep learning to forecast retinal disease from three-dimensional optical coherence tomography scans. The Heart and Vascular Institute, Cleveland, Ohio used a variation of the random forest technique to predict survival risk factors in patients with systolic heart failure. The University of Toronto published a comparison of machine learning methods to forecast gene function, and Zhejiang University School of Medicine has created a machine-learning algorithm to identify and prioritize cancer driver genes. This list goes on and on.
From such research, we can envision a future world in which condition-specific AI models trained on the world’s collective data are widely available, where genes are instantly and accurately interpreted at the click of a button, where a machine picks up the most subtle of clues in a scan with high precision, where certainty of diagnosis and prognosis are substantially higher.
It’s an appealing vision, and therefore, the excitement is understandable. However, the vast majority of machine learning activity remains in the sphere of research rather than clinical application. Data access, patient privacy obligations, nascent regulatory attention, and model explain ability must be carefully managed when bringing machine learning into practical use. The ethics and liability surrounding using algorithms to augment human judgment are substantial considerations, sharing much with the advancements in the field of autonomous vehicles. Even when healthcare safety matters are satisfactorily resolved, our collective responsibility to reduce energy consumption rears its head. Endless data and computing power may solve our healthcare woes, but at what energy cost to our planet.
In short, let’s remain excited and driven by the prospects of AI for healthcare but stay and patient and realistic about the complexities of implementation. It’s definitely going to be a long and exciting journey!