Those operating in the medical sector have to deal with strict laws and regulations. So did Nicholas. Therefore, training algorithms is difficult: medical data are not available, such as text for natural language processing models, for example. “The algorithm is trained on an anonymous data set, which is as heterogeneous as possible. We obtained this data collection through various collaborations at the national and international levels. Heterogeneous data sets are important here, because you want to prevent the algorithm from having bias and detecting certain groups of the population better than others,” explains Nizak.

But no matter how well an algorithm is trained, it is never 100% accurate. “In this case, false negatives are more annoying than false positives, because you don’t treat a person at all. So you’d rather detect too much and have your doctor double-check than miss something. But all the algorithms together – because it is not one algorithm that determines the diagnosis – has 90 to 95% accuracy.”

In addition, data sharing – between doctors and between assistants and doctors, all through apps – poses a challenge, because the data is very sensitive and therefore must be properly protected. For example, only people who really need access can access the data, and the data can only be used for a certain period of time.

The strict rules make it all the more striking that Nicolab runs its applications in the cloud, i.e. on AWS. “It was of course very scary for the hospital at first. So we develop apps from the ground up with security and privacy in mind, to comply with laws and regulations. So we make sure that doctors are properly authorized and you can see exactly who has access and when with a trailing audit. It is always transparent, both to the doctors and the patients whose data is involved.”