Evaluation of Different Learning Algorithms of Neural Networks for Drug Dosing Recommendations in Pediatrics

Publicly accessible databases with evidence-based information on drug dosages for children and adolescents are not available in Germany. In previous work a prototypical web-based online platform for pediatric dosing recommendation has been developed. Quality assured maintenance of such a database is a time consuming effort. Recent work has shown that it is possible to use routinely documented data for machine learning approaches in order to create models for future decision support tools. This work describes the development of a prototype for pediatric dosing recommendations on the basis of routine drug prescriptions. Since they are structured for daily clinical use, not for machine learning, they include a substantial proportion of narrative text that requires preprocessing with consideration of medical and pharmaceutical knowledge. Three different learning algorithms have been applied and compared. The genetic algorithm with backpropagation has achieved the highest accuracy in the predictions. Our study constitutes a first step towards pediatric dosing recommendations, but there are multiple additional steps to be taken before a routine use might be considered, such as an evaluation by experienced physicians.