Detection of pediatric developmental delay with machine learning technologies.

Area of Science:

• Pediatric Medicine
• Machine Learning in Healthcare
• Developmental Pediatrics

Background:

• Early identification of children with developmental delay (DD) is critical for improving outcomes through timely intervention.
• Current diagnostic methods can be costly and advanced, highlighting the need for accessible screening tools.
• Therapy visit frequencies represent a potentially low-cost data source for predicting DD.

Purpose of the Study:

• To investigate the efficacy of using therapy visit frequencies (physical, occupational, speech therapy) to predict developmental delay in children.
• To develop and evaluate machine learning models for predicting DD based on historical therapy data.
• To establish a cost-effective method for screening children at risk of developmental delay.

Main Methods:

• Utilized a dataset of 2,552 outpatients with 34,862 visits from a Taiwanese hospital (2012-2016).
• Developed and compared three machine learning models: Deep Neural Networks (DNN), Support Vector Machines (SVM), and Decision Trees (DT).
• Evaluated model performance using the F1 score, sensitivity, and positive predictive value.

Main Results:

• Decision Tree (DT) models demonstrated superior performance, particularly when high sensitivity was prioritized.
• The best-performing DT model achieved a sensitivity of 0.902 and a positive predictive value of 0.723.
• DT models outperformed DNN and SVM models in predicting developmental delay based on therapy frequencies.

Conclusions:

• Therapy visit frequencies are valuable predictors of developmental delay in children.
• Machine learning models, especially Decision Trees, can effectively predict DD using this readily available data.
• These cost-effective predictive models hold significant potential for widespread clinical application in early DD screening and intervention.