717 - Machine learning with population data to improve prediction of risk of death until 5 years for infants admitted into neonatal intensive care units

Publication Number: 717.5907

Kok Joo Chan, The University of New South Wales, Petaling Jaya, Selangor, Malaysia; Mithilesh Dronavalli, University of Western Sydney, Sydney, New South Wales, Australia; Sithum Munasinghe, Western Sydney University, Campbelltown, New South Wales, Australia; Meredith C. Ward, Royal Hospital for Women, Sydney, New South Wales, Australia; Ju-Lee Oei, Royal Hospital for Women, Randwick, New South Wales, Australia; Barbara Bajuk, Retired, Marsfield, New South Wales, Australia; Kee Thai Yeo, KK Women's & Children's Hospital, Singapore, N/A, Singapore

Background: Infants in neonatal intensive care units (NICUs) are at increased risk of death even beyond hospital discharge but prognostication of risk is impeded by confounding patient factors and difficulties in accessing long-term information.

Objective: To use linked population data to conduct clustering analysis for identifying characteristics associated with increased risk of mortality in infants admitted to NICU until 5 years.

Design/Methods: Data for 28,652 infants admitted into 9 NICUs in New South Wales, Australia (NICUS registrants) between 2007-2016 were linked to death registries until 5 years of age. K-means clustering was performed with 113 infant characteristics. Similarity was measured with Gower’s distance. The automation process grouped infants according to similarity of perinatal characteristics and treatment. Incidence rate ratios (IRR) for death before discharge home (DBD), and death up to 5-years of age (DB5) were determined with Poisson regression.

Results: Eight distinct clusters were identified (Figure 1). Cluster 1, the comparator group, was the largest (n=17,501) and comprised mostly of term [median gestation 37 (IQR 34-39) weeks] infants with NICU length of stay of 8.7 (SD ±11.1) days, and low DBD (0.44%) and DB5 (1.86%) risks.

Prior to hospital discharge, clusters with the highest risk of death were Clusters 8 (IRR DBD: 22.0, 95%CI: 12.9-37.8) and 6 (IRR DBD: 14.7, 95%CI: 10.8-20.0), comprising preterm infants, median gestation 25 and 27 weeks, respectively. Next were term infants with specific medical issues, including Cluster 5 (IRR DBD: 13.3, 95%Cl: 7.0-25.7) with prenatal opioid exposure (13.5%) and severe hypoxic ischemic encephalopathy (25.2%), and Cluster 7 (IRR DBD: 5.2, 95%CI: 3.1-8.6) with maternal gestational diabetes (30.2%) and major cardiac anomalies (36.8%).

By 5-years of age, Cluster 8 (IRR DB5: 6.6, 95%CI: 4.1-10.4) continued to have the highest risk of death but this was followed by Cluster 7 (IRR DB5 :5.6, 95%CI: 4.2-7.5) and Cluster 6 (DB5: 4.1, 95%CI: 3.2-5.1). At 5 years, Cluster 4 (IRR DB5: 3.2, 95%CI: 1.8-5.6), comprising of late preterm infants (median gestation 36 weeks) who were more likely to be male (78.8%), indigenous (37.2%), exposed to maternal smoking (58.4%) and born in non-tertiary hospital (68.0%), also had increased risk of death.

Conclusion(s): Extremely preterm infants had the highest risks of death before and after hospital discharge to 5 years of age. Certain groups of term and late preterm infants, however, were also at higher risk of death, which warrants consideration of prognostication and follow-up support services.

Figure 1
Forest plots demonstrating comparison of risks of DBD and DB5 between clusters with Cluster 1 as reference

Table 1
Examples of infant demographics and characteristics used for clustering

Table 2
Risks of death before discharge home and death up to 5-years of age based on clusters