Machine Learning powered Patient Records Analysis for Injury Monitoring in Children and Adolescents

Abstract Background Injuries to children and adolescents are common and are associated with individual suffering and costs. Switzerland lacks data of injuries in children and adolescents. This study investigates the feasibility of utilizing machine learning (ML) methodology to automatically extract relevant data from electronic pediatric patient records needed for monitoring and prevention purposes. Methods The feasibility study follows three steps. 1. Utility evaluation of text data from electronic pediatric patient records regarding ML, injury monitoring and prevention requirements in a test-sample. 2. The development of an ML-based approach to extract relevant information based on electronic pediatric patient records (N = 30’884) of children treated for injuries at the University Children’s Hospital Zurich emergency department between 2018-2022. 3. The performance of the ML methodology is evaluated. Results Qualitative and ML expert assessment of the sample showed that the desired data on injuries is contained in the test-sample and necessary requirements are met. Data records differ in detail with severe injuries providing especially rich information records. The international IDB code-tree was proved to be too elaborate and adult-orientated and was adapted. Preliminary data analyses indicate that the number of accidents is highest in the age group from 0 to 4 years (36%). More male (58%) children were treated. Majority of accidents were classified as less severe (73%), few (<1%) as very severe. Further lessons learned, along with more detailed prevalence data on injuries in children and adolescents, will be available by August 2024. Conclusions Extraction of injury information from electronic patient record by means of ML may provide valuable information for injury monitoring and prevention, so far missing in Switzerland. Prerequisites for the successful application of ML are digital record access, usable coding-tree, and detailed text input by the emergency departments. Key messages • An ML-based approach has the potential to improve the data basis by extracting valuable accidents information from existing texts of patient records. • To successfully apply this method, digital record access and detailed text data are needed.