159 - Risk Stratification using Machine Learning to Inform High-Risk Infant Follow-Up Enrollment of Term and Late Preterm Infants

Poster Number: 159
Publication Number: 159.363

Katherine A. Carlton, Medical College of Wisconsin, Mequon, WI, United States; Jian Zhang, Medical College of Wisconsin, Milwaukee, WI, United States; Ke Yan, Medical College of Wisconsin, Milwaukee, WI, United States; Samuel J. Adams, Children's Hospital of Wisconsin, Milwaukee, WI, United States; Erwin T. Cabacungan, Medical College of Wisconsin, Wauwatosa, WI, United States; Susan S. Cohen, Medical College of Wisconsin, Milwaukee, WI, United States

Background:
Infants ≥ 34 weeks gestation are not routinely referred to high-risk infant follow-up (HRIF) programs at neonatal intensive care unit (NICU) discharge. AAP guidelines for hospital discharge of the high-risk neonate define this high-risk population as: (1) preterm, (2) those with special health care needs or dependent on technology, (3) those at risk because of family issues, and (4) those with anticipated early death. It is unclear if these categorizes adequately identify which term and late preterm infants benefit most from HRIF. Further analysis is needed to identify key NICU clinical factors that can be used to risk-stratify infants for HRIF enrollment.


Objective: To identify a risk stratification model for HRIF enrollment of infants ≥ 34 weeks gestation using machine learning. We hypothesize that specific clinical factors better predict abnormal developmental screening in early childhood.

Design/Methods:
We performed a retrospective cohort study of infants born ≥ 34 weeks gestation between 2013-2015, admitted to our level 4 NICU, and received developmental screening or formal neuropsychology testing until age 6 years. Assessments were considered abnormal if concerns in cognition, motor, language, and/or social domains were documented. Machine learning by CART (classification and regression tree, CART® engines, Salford Predictive modeler software) analysis was performed using 46 clinical factors to predict infants with abnormal developmental screening in early childhood. Clinical factor categories included: demographics, delivery interventions, diagnoses, procedures, nutrition, and discharge planning.


Results:
We screened 1355 NICU infant charts, and 1183 (87%) infants received developmental screening (Figure 1). We identified 557 (47%) infants with abnormal developmental screening results. Demographic characteristics of the study cohort are displayed in Table 1. Results of machine learning are demonstrated in Figure 2. Algorithm sensitivity and specificity were 67% and 75%, respectively. The area under the receiver operating characteristic curve was 0.73.


Conclusion(s):
Our machine learning algorithm risk-stratified infants with ≥ 3 medications or infants with ≤ 2 medications and ≥3 follow-up appointments as high-risk for abnormal developmental screening by 6 years of age. We speculate that these NICU clinical factors can be used to risk-stratify infants in real-time for HRIF enrollment prior to NICU discharge.