Artificial intelligence in medicine - where are we at?

By guest author: Dr. Isabella Hermann - Ms. AI Ambassador for AI & Politics

The outbreak of COVID-19 has triggered considerable coverage of how artificial intelligence (AI) can combat the spread of the pandemic. AI has a role to play, for example to speed up the development of a vaccine, but overall, the effect of the technology seems rather small. But what can AI actually do when it comes to medicine? Wouldn't it be a dream – like in many science-fiction-films – to place our sick or wounded self in the care of a medical device that gives us the right treatment automatically? In science-fiction stories, such "autonomous medical pods" are a popular motif for illustrating scientific progress in medicine. Right now, there are some medical applications that might make us think that we are not far from inventing such devices. But a closer look quickly reveals how limited current systems still are - and how susceptible to bias.

Let’s first take a look at the field of fiction. In the film "Elysium" playing in 2154, it takes just ten seconds from the detection of a fatal leukemia by an autonomous system to healing, after which the patient instantly opens her eyes again. In "Passengers" which is located in the year 2343, the severely injured and already clinically dead Sam Worthington is brought back to the living in a medical pod and can immediately kiss his companion Aurora alias Jennifer Lawrence passionately. And in "Prometheus", the prequel to the famous "Alien" films and set in 2089, the protagonist has the growing alien embryo in her stomach cut out in an autonomous pod, is fully automatically sewn up again and can continue to fight her way through the spaceship.

How far away from this is the current state of research? In the United Arab Emirates, for example, so-called "health pods", which can be found in shopping malls since 2018, should help diagnose health problems in less than ten minutes. The individual pods have a built-in full body scanner that measures vital functions such as blood sugar, height and weight and classifies them with the help of algorithms. This can provide initial orientation, but of course does not replace medical treatment.

AI is strong where it can work with images

Currently, AI is driving developments in medicine. Apparently, AI-based systems perform well in areas where images are used to make diagnoses. These are mainly radiological procedures that are used, for example, in cancer detection. For this purpose, an AI system is trained with labelled data, for example cancerous and non-cancerous tissue, until it recognizes certain patterns and can thus distinguish diseased from healthy tissue.

Such diagnostic systems can now be applied quite reliably to support physicians and are already being used in hospitals. However, they are only as good as their data. One example: If an AI for skin cancer detection has been trained mainly with images of patients with light skin color, it will not work for people with darker skin color. Human error - including discrimination - can thus be transferred into the systems. In general, the systems can easily be irritated if the characteristics of the actual persons using it differ from those in the training data set. In this case, characteristics can lead to false diagnoses that physicians sometimes cannot even follow: the "black box" problem of AI systems.

Autonomous surgical robots are still dreams of the future

When it comes to surgical robots we are even further away from autonomy. Until now, surgical robots have been used, if at all, only at a very low level of autonomy with the surgeon still retaining full control. It is up to debate whether these robotic systems even benefit the patient as well as the surgeon and save costs for the health care system. Actual benefits would probably only arise if the machines were able to operate more independently. But it is precisely these higher levels of autonomy that are difficult to implement. On the one hand, a robot must be able to work in a context-based manner and also be able to cope with unforeseen situations, for example if your stomach is suddenly full of blood during an abdominal operation. On the other hand, autonomous systems could be particularly useful where human surgeons reach their limits. This is the case, for example, when the surgery requires a great deal of skill or involve areas that are difficult to reach. In both cases, however, the data needed to train the robots sufficiently is naturally lacking. At the moment, a surgical robot can at best learn what a human surgeon can do, but not what a human surgeon cannot do.

So, we are still a long way from "Autonomous Medical Pods", which could offer diagnosis and therapy or even surgery from a single device. Nevertheless, as technology advances, we have to deal with fundamental questions: How will doctors be ethically and legally responsible in the face of ever-increasing automation? And what does this mean for the doctor-patient relationship? Because in the end one thing is clear: machines may optimize, but they are not empathetic and caring.