(333) Generalizing NICU Admission Predictions for Medicaid Patients Using Machine Learning

Medicaid patients have higher rates of maternal and neonatal complications. No models have been designed to specifically examine this patient population. We evaluated the generalizability of a NICU admission prediction model trained on a mixed population of insured patients (commercial, Medicaid, self-pay) when applied specifically to Medicaid patients.

Study Design:

An extreme gradient boosting model was trained on data from years 2014-2021 of the Vital Statistics System of mixed insurance coverage with the exclusion of year 2015 due to unavailable data. The model was tested on a mixed insurance and solely Medicaid-covered cohort. Inclusion criteria for the Medicaid testing set were deliveries covered by Medicaid in the year 2022 with available NICU admission data. Inclusion for mixed testing cohort were 2022 deliveries with available NICU admission data. Cases with missing NICU admission data and those from hospitals not reporting NICU admissions were excluded. The primary outcome of interest was NICU admission immediately post-delivery. 

Results:

38 clinical variables for 15,782,488 obstetric patients were included in the training data. The testing set consisted of 1,500,152 Medicaid patients, with 157,871 (10.5%) experiencing NICU admissions. The model demonstrated a strong performance on the Medicaid cohort with an AUC of 0.73, 80.3% accuracy, and a 0.80 F1 score. 

When tested on the 2022 mixed cohort (self pay, commercially insured, Medicaid, and other) of 3,666,784 patients, of which 347,586 (9.5%) patients had a NICU admission, the model had an AUC of 0.73, 83.3% accuracy, and F1 score of 0.83. The analysis revealed that the most influential factor in predicting NICU admissions was the interval since the last live birth.

Conclusion: The NICU prediction model, initially trained on a mixed population of insured patients, demonstrated strong generalizability and robust performance when applied to Medicaid patients in 2022. By leveraging routinely collected data, the model can aid in reducing NICU admissions and associated healthcare costs, ultimately improving outcomes for Medicaid patients.