By Deborah Borfitz 

May 27, 2021 | A machine learning specialist at Aalto University (Finland) is one of the lead developers of an algorithm for determining drug dosing in children at various stages of maturation, which could one day aid in the design of pediatric clinical trials having a small number of participants. Methods for modeling developmental and physiological differences between children and adults are “quite limited” and, in some cases, can lead to false conclusions because there is too little data to feed into the calculation, according to Eero Siivola, a doctoral candidate in advanced exploratory analytics in the school’s department of computer science.

During his recent internship at Novartis (Switzerland), Siivola worked on the model that uses Gaussian process regression—a popular computational approach for detecting deviations when small datasets are being used to make predictions. The same method is being widely applied in different disciplines to, for instance, predict ore concentration in the ground or the volatility of financial markets. 

It is likely the first time Gaussian process regression has been used to model how the organs of small children develop, Siivola adds. The technique involves making virtual observations, essentially “safe assumptions” that get formulated into data inputs, such as the organs of older teens are closer to that of adults and the organs of smaller kids change faster. 

The novel method, described in an article recently published in Statistics in Medicine (DOI: 10.1002/sim.8907), makes better use of available data and could help determine safe drug doses more quickly and with fewer per-patient observations, says Siivola. By re-analyzing a pediatric trial investigating everolimus, a Novartis drug used to prevent the rejection of organ transplants, researchers demonstrated the model can reliably detect maturation trends from sparse pediatric data. 

The model comes out of a research program at the Finnish Center for Artificial Intelligence on agile and probabilistic AI and can be used to simulate the effects of different dosing levels in children, based on a drug’s concentration level in different parts of their body as a function of time, Siivola explains. The method “doesn’t make very specific assumptions [and]… takes specific care in modeling uncertainty related to the predictions,” but can be applied to a wide variety of drugs being newly tested or whose indications are being extended. 

“The hope is that people will use the open-source code to experiment with their own datasets and realize that the methodology can be as good and even better than the methods they are [currently] using,” says Siivola. With sufficient uptake, the machine learning approach could eventually be used in registered clinical trials—pediatric or otherwise. 

The new model could potentially “make the existing but vague maturation models more principled,” says Siivola. In the clinical trial arena today, drugs are generally tested first in adults and the results are used to select the first doses administered to pediatric study participants. 

Even in the clinic, pediatricians will often use blood or urine tests to cautiously watch for a child’s initial response to a medicine with a narrow concentration window, Siivola says. Most of the drugs they are prescribing have not been specifically licensed for the treatment of children and, if they have, small sample sizes can limit the generalizability of findings.

In addition to the relatively small number of available participants, he continues, regulatory agencies have provided additional protections for children involved as subjects in research so their enrollment is often less than straightforward. Parental consent (and sometimes also an older child’s assent) is required and there may also be technical challenges related to the ethics of testing frequency or how well children can express how they are feeling. 

Understanding how a drug works for all kids based on a pediatric trial enrolling 20 participants is problematic because it requires a lot of guesswork, says Siivola. Variability in body size among children is also “much larger” than in adults, and the organs undergo many changes related to growing up. “Trying to understand these very complex effects from a very limited sample size increases uncertainty.”

Up to now, efforts to close the knowledge gap on pharmacokinetics in children have relied on parametric functions of age as a proxy for maturation, says Siivola, who is working on his Ph.D. in applications of the Gaussian process technique. Existing parametric models “do not necessarily require more data, but they are [sometimes] bad at modeling uncertainties related to predictions.” Consequently, researchers might incorrectly interpret results that can influence dosing levels for some age groups.

As a condition of his internship at Novartis, results of his basic research on the pediatric drug-dosing algorithm were published as an open-access article under the terms of the Creative Commons Attribution License, Siivola says. The source code is openly available to the larger scientific community.

Sebastian Weber, co-author on the Statistics in Medicine paper and a statistical methodologist at Novartis, continues to work on quantitative methods to optimize the design of randomized trials, Siivola says. Novartis declined to comment about any ongoing work with the new machine learning model to assess drug dosing.