Utility of medical record diagnostic codes to ascertain attention-deficit/hyperactivity disorder and learning disabilities in populations of children

Abstract

Abstract Background To develop and evaluate machine learning algorithms to ascertain attention-deficit/hyperactivity (ADHD) and learning disability (LD) using diagnostic codes in the medical record. Method Diagnoses of ADHD and LD were confirmed in cohorts of children in Olmsted County of Minnesota based on validated research criteria. Models to predict ADHD and LD were developed using ICD-9 codes in a derivation cohort of 1057 children before evaluated in a validation cohort of 536 children. Results The ENET-MIN model using selected ICD-9 codes at prior probability of 0.25 has a sensitivity of 0.76, PPV of 0.85, specificity of 0.98, and NPV of 0.97 in the validation cohort. However, it does not offer significant advantage over a model using a single ICD-9 code of 314.X, which shows sensitivity of 0.81, PPV of 0.83, specificity of 0.98, and NPV of 0.97. None of the models developed for LD performed well in the validation cohort. Conclusions It is feasible to utilize diagnostic codes to ascertain cases of ADHD in a population of children. Machine learning approaches do not have advantage compared with simply using a single family of diagnostic codes for ADHD. The use of medical record diagnostic codes is not feasible to ascertain LD.

Keywords

• Attention-deficit/hyperactivity disorder
• Learning disability
• Diagnostic code
• Machine learning
• Validation