Development and Validation of Prediction Model for Neonatal Intensive Care Unit (NICU) Admission Using Machine...

The work aims to provide a method for assessing newborn infants' risk to determine whether they need to be admitted to the NICU before birth. Term neonates who were admitted to the NICU are known to have health problems, a greater mortality risk, and higher healthcare costs. Researchers want to develop an algorithm that estimates the risk of NICU admission for this particular subset of newborns using an integrated statistical approach. This risk assessment might lower the morbidity, mortality, and healthcare expenses related to NICU hospitalizations by assisting in the early identification of possible instances.

The data was collected from a multispecialty hospital using hospital-based records from the obstetrics and gynecology department. All the clinical and demographic parameters are described as per requirement. A multivariate statistical analysis was done to identify potential risk factors for NICU admission. Four classification models were used to predict NICU admission. All the models were evaluated based on their performance matrices.

In multivariate analysis, we found Preterm deliveries (β = 1.003 Aor = 2.727 95% CI = 1.54,4.80 P < 0.001), Hypertension (β = - 1.419 Aor = 0.242 95% CI = 0.112,0.523 P < 0.001), AFI (β = 1.262 Aor = 3.53 95% CI = 1.06,11.69 P = 0.039), Birth weight (< 2.5 kg) (β = 1.011 Aor = 2.75 95% CI = 1.57,4.81 P < 0.001), Mode of Delivery(LSCS)(β = 1.196 Aor = 3.307 95% CI = 1.95,5.60 P < 0.001) and maternal complication (β = 6.962 OR = 7.69 95% CI = 5.67,13.69 P < 0.001) are the potential risk factors for NICU admission. The decision tree performed the highest accuracy (0.921) and AUC (0.966) as compared to other models to predict NICU admission.

Using an explainable feature learning technique to predict NICU admissions contributes to better global health data utilization and a more hopeful future in healthcare.

The online version contains supplementary material available at 10.1007/s13224-024-02009-0.