הכנס ה-24 של העמותה להתפתחות הילד ושיקומו

A natural language processing pipeline for Gross Motor Function Classification System level prediction

Kelly Greve 1 Yizaho Ni 2 Brad Kurowski 3 Amy Bailes 1 Jilda Vargus-Adams 3 Mary McMahon 3 Bruce Aronow 2 Alexis Mitelpunkt 4
1Occupational Therapy and Physical Therapy, Cincinnati Children’s Hospital Medical Center, USA
2Biomedical Informatics, Cincinnati Children’s Hospital Medical Center, USA
3Pediatric Rehabilitation Medicine, Cincinnati Children’s Hospital Medical Center, USA
4Pediatric Rehabilitation, Dana Children's Hospital, Israel

Introduction: A critical part of children’s with Cerebral palsy (CP) care is characterization of function over time. Gross Motor Function Classification System (GMFCS) is a commonly accepted way to classify mobility severity for individuals with CP but it is not documented consistently in clinical practice. A unique way to standardize tracking of function is application of Natural Language Processing (NLP) to the electronic health record (EHR). NLP is an area of artificial intelligence focused on training computer programs to understand written human language in a way that a computer comprehends.

Aim: To predict ambulatory status and GMFCS Levels in patients with CP applying NLP to EHR clinical notes.

Materials and Methods: Retrospective cohort divided with stratified random sampling into model training (80%) and test/validation sets (20%). The study included 2,621 patients with 789,246 encounters and 1,993,405 clinical notes from the EHR at a tertiary pediatric medical center. More children were male (56%), White (75%). Notes containing GMFCS documentation were excluded in the model creation. Clinical notes were preprocessed, n-gram features (n≤2) extracted and weighed. Two machine learning classifiers implemented targeting outcome for each encounter formatted as a binary-class. The predictive models with optimal parameters were applied to the validation group.

Results: The median age was 15 (IQR 10 – 20) years for the total cohort. The validation group resulted in 70% SEN, 88% SPEC, 81% PPV, and 0.89 AUC for predicting ambulatory status. NLP applied to the EHR differentiated between GMFCS Levels I-II and III (15% SEN, 96% SPEC, 46% PPV, and 0.71 AUC); and IV and V (81% SEN, 51% SPEC, 70% PPV, and 0.75 AUC).

Conclusions: NLP is an effective approach for determining GMFCS level in children. Potential clinical applications include identification of children at risk for functional deficits who may require further evaluations based on clinical notes.