188 - Identification of Risk Factors for Prolonged Length of Stay in Very Low Birth Weight Preterm Infants: An Explainable Machine Learning Approach

We used machine learning (ML) to identify risk factors for prolonged LOS in VLBW preterm infants.

Design/Methods:

This retrospective study collected data on VLBW preterm infants born between 2016 and 2019 from the Taiwan Neonatal Network. The prolonged LOS group was defined as the 25% of infants with the longest LOS in each gestational age category. We used 38 variables and constructed 5 ML models by using Waikato Environment for Knowledge Analysis (Weka) and the Orange Data Mining software. The permutation feature importance technique was used to interpret the importance of each risk factor. The CfsSubsetEval with Best First method in Weka was used for attribute (factor) selection. The performance of the ML models was compared using the receiver operating characteristic curve area under the curve (AUC).

Results:

A total of 3581 VLBW preterm infants who were discharged alive were enrolled. The gradient boosting (GB) model had the best AUC (0.77). We selected 7 factors, namely birth body weight, inhaled nitric oxide use, bronchopulmonary dysplasia (BPD) status, steroid use for BPD, surgery for necrotizing enterocolitis (NEC), surgery for conditions other than PDA or NEC, and enteral feeding type. The performance of the simplified models using 7 factors was not inferior to that of the models using all 38 considered factors (AUC 0.77 vs 0.758 in the GB model). BPD status was the most crucial factor for prolonged LOS.

Conclusion(s):

We identified 7 factors associated with prolonged LOS in VLBW preterm infants in this population-based study. BPD status was the most crucial factor for prolonged LOS.