Pediatric cancer is not only rare, there are many different types. And within those types, there are also many different subtypes. A precise diagnosis is important to be able to give each child the most appropriate treatment. Therefore, innovative diagnostics is also one of the components of the Máxima Center's multi-year strategy.
The pathologists and hematologists of the pediatric oncology laboratory establish the diagnosis by examining the composition of tumors very precisely. In doing so, they also use information from molecular analyses, such as the RNA analysis performed for every child under treatment at the Máxima. However, current analyses look at only a very limited portion of these RNA data.
Fleur Wallis, PhD candidate with associate research group leader Dr. Lennart Kester and research group leader Dr. Patrick Kemmeren, developed an AI algorithm that automatically determines the tumor type from the large amount of data resulting from an RNA analysis. The pathologist or hematologist uses the classifier as a tool. For example, to confirm their own diagnosis or, in case of doubt, to still be able to make the exact diagnosis. In doing so, the algorithm in some cases pinpoints subtypes that cannot be determined by manual examination.
Today, the results and the downloadable algorithm were published in eBioMedicine. This research was made possible thanks to KiKa and a financial contribution from the Adessium Foundation to the data infrastructure.
Pediatric cancer-wide
M&M is the first pediatric pan-cancer algorithm and covers 52 different tumor types and the 96 related subtypes. To train the algorithm, tumor material from 1256 children was used. How to train an algorithm explains Wallis: 'We look in the RNA data for patterns that best show the difference between different tumor types. Then we train the model 100 times, each time hiding a different part of the data. For a child with an unknown diagnosis, all 100 models predict what type of tumor it is. We combine these predictions into one definitive answer. By counting how many of the 100 models give the same prediction, we show how certain the model is.'
After checks, the algorithm turns out to identify the correct tumor type in 99% of cases and also the correct tumor subtype in 96% of cases. This percentage remained virtually the same when the algorithm was also used with data collected outside the Máxima.
Further research
Diagnosis is one of the applications of the AI algorithm. Kester also sees opportunities for research into the origins of childhood cancer. 'A next step is to investigate why the algorithm makes certain choices, which parts of the RNA data patterns are now specific to a certain tumor type? And can we link this, for example, to processes that cause tumor cell growth? We could still learn a lot this way about the connection between what we see in the RNA and the biology of cancer.'
In the future, Kester also hopes to start linking this to the child's treatment and its effectiveness. To do that, more data must be collected in the coming years. In the Máxima, but also in collaboration with other pediatric oncology centers. Kemmeren adds: 'It is essential that the data we need is not only accessible, but that we can also easily apply the algorithms we develop in the various centers.' Kester: 'Ultimately, in this way, we hope to use data and the AI algorithm to further improve diagnostics so that we can provide the most appropriate treatment to every child with cancer.'