They examine hundreds of x-rays, scarcely ever overlooking a suspicious detail. Their decisions are consistent. And they identify correlations even in vast volumes of complex medical data: Artificially intelligent diagnostics systems have the potential to revolutionize the identification of diseases. Simply because they are capable of doing things that neither humans nor other technologies have been able to do until now. The consequences: more highly targeted treatments and a reduction of unnecessary costs and patient risks.

How reliable are the systems?

Nevertheless, for the latter, the possibility of increasing the reliability is growing: by means of “Explainable AI” (XAI). This young branch of AI research aims to make the decisions of algorithms more comprehensible to humans. Viktor Kölzer and his team are also working with XAI. Another way to verify the reliability of algorithms is to validate them using as many different datasets as possible. Viktor Kölzer: “The more data that is available for this purpose worldwide, the better. Without sufficient – digitally recorded – data and the opportunity for international exchange, we will not be able to make further progress in the field of AI.”

Major challenge: data privacy

So let’s now address the pivotal challenge relating to AI: data privacy. “The existing Swiss Data Protection Act is definitely not yet optimally tailored to the new applications,” says Markus Christen. And he immediately follows up with a possible approach towards solving this issue: “The data from one sphere of life should not be allowed to leave its confines.” As a rule, people who have agreed to the use of their data in pathology would generally not object to it also being used in oncology. “However, they would rightly object if, all of a sudden, their data is also used by insurance companies or banks,” Markus Christen explains. “If instead of the minor boundaries – between individual fields of research or projects – the major boundaries were defended, medicine would suddenly gain access to much more data.”

An even more typical ethical problem related to AI: How representative is the data that was used to train the algorithm? In principle, according to both Markus Christen and Viktor Kölzer, algorithms have the potential to make decisions that are much more nondiscriminatory than, for example, a small team of physicians. This, however, only applies if the data used for training also adequately reflects, for example, all the relevant patient groups: “The more unbiased the data, the more unbiased the decisions made by the intelligent system.”

Huge step in developing countries

This also means that an AI system cannot simply be transferred to another part of the world. For example, the intelligent procedure from Oxford University, which reputedly outperforms British cardiologists in the diagnosis of heart disease, will not necessarily achieve the same results in Malawi. On the one hand because of the training data: For the system to perform equally as well, it would have to be trained using local patient data. But other technical requirements also play a role: simply using a different scanner can irritate the AI. “The AI can learn such differences and take them into account,” Markus Christen explains. “If it fails to do so, its results will be skewed.”

Especially in diagnostics, algorithms can provide people with access to methods that might otherwise be completely missing

When all these factors are taken into account, AI systems mean one thing above all for developing countries: a giant leap forward. “In certain regions of Africa, for example, there are only very few specialists for each medical field,” Viktor Kölzer emphasizes. “Especially in diagnostics, algorithms can provide people with access to methods that might otherwise be completely missing.” Such methods can be made available to hospitals or medical practices – or, thanks to the widespread use of smartphones, directly to the patients. For example in the form of apps that analyze moles or measure blood sugar and heart rate.

Inexpensive access to modern cancer diagnostics

At least in our part of the world, the colon cancer algorithm developed by Viktor Kölzer’s team – which is currently being tested to determine whether it can be generalized – could soon constitute a major step forward. “If the system proves successful, we will be able to make it available at very low cost – for five to ten Swiss francs per patient,” Viktor Kölzer explains. In comparison: Conventional RNA sequencing – which, albeit, also provides important additional information for cancer diagnostics – costs around 1000 Swiss francs per patient.