Julia Vogt, how do we benefit from doctors using AI?

AI is good at handling large amounts of data and can detect irregularities in images that might escape the human eye. If the AI highlights where something looks different, doctors can focus their attention on those areas and make their own assessment. This saves time without compromising medical quality. 

In my research group, as an example, we developed an AI model that helps to diagnose heart defects in newborns. During the baby¡¯s ultrasound examination, the model analyses the image data and flags any deviations from the normal pattern. When used in this way, AI becomes a valuable tool for physicians and is far from being a mysterious black box that spits out opaque results. 

In another research project, we developed an early detection system for neonatal jaundice. Today, many mothers and their babies leave the hospital just one day after the baby is born, but jaundice often only develops later. Our model can reliably predict the risk of jaundice based on four markers: the baby¡¯s age and weight, the gestational week at birth and the bilirubin level in their blood. If our model indicates a higher risk, the baby can remain under observation in the hospital longer or the parents can receive targeted guidance. 

This example illustrates AI's potential in early detection. By combining complex data with specific markers, AI models can identify patterns that indicate early stages of an illness before any symptoms are noticeable. The potential is particularly high when these markers come from different sources, such as images, lab results, and genetics. AI is able to integrate heterogeneous datasets, present them in a way that is easy to understand and analyse them holistically. Physicians can then interpret this analysis and decide on the best course of action. 

In another research project focusing on appendicitis in adolescents, we developed a web interface where physicians can upload ultrasound images, tabular clinical data or clinical notes on body temperature and pain levels. The AI model then analyses whether appendicitis is present and whether complications are likely. If applicable, it also recommends a surgical intervention. In doing so, the model provides a transparent explanation of how it arrives at its conclusions. 

The AI model was trained on data from one single hospital. We are now working to generalize the model so that it can be applied more broadly, rather than just under our local conditions. By validating with additional datasets, we consistently verify that the model does not draw any false conclusions. Of course, no one can guarantee a complete error-free AI, just as humans can make mistakes. But one advantage that AI will always have over us? It never gets tired.¡±