Hospital readmissions negatively impact patient quality of life and incur substantial healthcare costs. To target resources to prevent readmission, this study sought to develop clinical prediction models for 30-day readmission in the general pediatric population and for children with medical complexity (CMC).

Sociodemographic and clinical characteristics were extracted from electronic health records for pediatric patients aged 6 months to 18 years admitted at an urban academic medical center between October 1, 2010 and July 31, 2016. Factors associated with unplanned 30-day readmission on univariate screen were candidates for the multivariable logistic regression models. Using backward selection, we derived a model predicting readmission utilizing characteristics obtainable at admission ("model at admission"). A second model was derived including variables available by hospital discharge ("model at discharge"). Model performance was assessed using c-statistic and calibration curves, and bootstrap resampling was performed for internal validation. CMC-specific models were developed and evaluated by repeating these procedures in the subgroup of medically complex children.

Of the 7,068 general pediatric index admissions during the study, 313 (4.4%) had an unplanned readmission within 30 days. The model at admission included the following variables: non-English language, prior admissions, prior emergency department (ED) visits, number of home medications, medical complexity, technology assistance, and medical versus surgical admission (c-statistic 0.68). The model at discharge included all these variables plus length of stay, weekday discharge, and discharge disposition (c-statistic 0.69). For the CMC subgroup, of 2,296 index admissions, 188 (8.2%) had readmissions. The CMC model at admission included prior admissions, prior ED visits, number of complex chronic conditions and medical versus surgical admission (c-statistic 0.65). When including variables available at discharge, the model also included length of stay, weekday discharge, and discharge disposition (c-statistic 0.67). Patients in the highest risk quartiles had 3.6 to 4.5 times higher readmission rates compared with patients in the lowest risk quartiles for all models. Bootstrap samples had similar c-statistics, and slopes did not suggest substantial overfitting in any model.

In conclusion, easily obtainable clinical characteristics are useful in identifying children at particularly high risk for readmission. These high risk children may be an appropriate target for interventions to prevent readmissions. Future proposals will involve external validation of the models and will explore whether the models can be used to target resources aimed at decreasing readmissions.