Development of an Electronic Health Records-based Predictive Model to Identify Children at Risk of Recurrent Acute Otitis Media
Jillian H Hurst, Congwen Zhao, Eileen M Raynor, Christopher W Woods, Matthew S Kelly, Michael J Smith, Benjamin A GoldsteinAbstract
Background
Recurrent acute otitis media (rAOM) affects 10-15% of children in the United States and is a leading cause of healthcare utilization and antibiotic prescriptions, accounting for over $1.3 billion in healthcare costs annually. Current treatments for rAOM include expectant management and tympanostomy tube (TT) placement; however, recent studies have suggested limited benefits of TT placement. Prospective identification of children at greatest risk of rAOM could help target interventions such as TT placement and identify new risk factors to guide preventive approaches. We sought to develop predictive models to prospectively identify children at risk of rAOM using electronic health records (EHR) data.
Methods
We extracted retrospective EHR data for children who were born in the Duke University Health System between January 1, 2014 and June 30, 2022 who had well child visits from birth through the fourth year of life, and who had at least one AOM episode during the study period. We identified all children in the cohort who met criteria for rAOM, defined as 3 or more AOM episodes within 6months or 4 or more episodes within 12 months. We used LASSO to fit two types of models to predict rAOM: an overall model, which did not account for episode order, and episode-specific models that only used data corresponding to a particular episode.
Results
Of 6,566 eligible children, 1,634 (25%) met criteria for rAOM. Children with rAOM and those who only had sporadic AOM episodes differed by sex, race/ethnicity, birth mode, age at first AOM episode, and outpatient healthcare utilization, as evaluated by standardized mean difference (SMD; Table 1). Predictive models using data from all episodes performed moderately well, as measured by the area under the receiver operating curve (AUC (95% confidence interval)): 0.81 (0.79, 0.83)). A model using data from birth to the time of the first episode had an AUC (95% CI) of 0.75 (0.72, 0.77); model performance improved with each subsequent episode (Table 2). Features associated with rAOM development included younger age at the time of first episode, number of prior antibiotic prescriptions, Hib doses, and outpatient encounters.
Conclusions
Our findings indicate that clinical exposures, healthcare utilization, and comorbidities documented in the EHR distinguish children who are at risk of developing rAOM. This model can be used to identify children for early referral to specialty care and to help guide future investigations of factors that influence rAOM development.