267 - Logistic Regression Compared to Machine Learning in Prediction of Neurodevelopmental Outcomes in Preterm Infants: A Population-based Study

Poster Number: 267
Publication Number: 267.225

Jehier Afifi, Dalhousie University Faculty of Medicine, Halifax, NS, Canada; Samuel Stewart, Dalhousie University, Halifax, NS, Canada; Tahani Ahmad, Dalhousie University, Halifax, NS, Canada

Background: Preterm infants are at high risk of mortality and morbidity including neurodevelopmental impairment (NDI). Accurate and early identification of those infants at risk for NDI enables early referral to targeted intervention programs with the potential to improve their functional development and quality of life. Traditionally, logistic regression (LR) has been used to develop prediction models of NDI in neonatal literature. Little is known about the role of machine learning in prediction of NDI in preterm infants, using clinical predictors that can be readily abstracted from patient records.  

Objective: To compare the accuracy of conventional LR to machine learning methods in prediction of NDI at 36 months of corrected age in very preterm infants ( < 31 weeks’ gestation).

Design/Methods: This population-based study used the AC Allen Provincial database to develop prediction models of NDI in a cohort of very preterm infants (220-306 weeks’ gestation) born in Nova Scotia, Canada between 2004-2016. NDI was defined as any of: cerebral palsy, Bayley scores < 85 in any domain, hearing or vision impairment). Prediction models were developed using four algorithms: LR with stepwise variable selection, elastic net regression, random forest (RF) and gradient boosting (GB). Models were compared using Area under the Curve (AUC) and their diagnostic properties were computed. The population sample was randomly split into training and testing datasets using a 70:30 ratio for development and validation of the prediction models, respectively, stratified on outcome.

Results: In this cohort, 197/665 children (30%) developed NDI as shown in Figure 1. Table 1 shows the prenatal, perinatal and postnatal characteristics of those children with and without NDI. On internal validation, all models including LR, elastic net regression, RF and GB provided good discrimination of children with and without NDI with comparable predictive accuracy (AUC ranged from 0.70, 95% CI: [0.60, 0.79] for RF to 0.74, 95% CI: [0.65, 0.83] for elastic net regression) [Table 2].Conclusion(s): Using clinical predictors, elastic net regression produced the most accurate model of the four algorithms selected in prediction of NDI in a population-based cohort of very preterm children, though there was minimal differentiation between the four algorithms. All methods can be used for predicting NDI in preterm children.
Figure 1: Flow Chart of a Population Cohort of Very Preterm Infants in Nova ScotiaAbbreviations: NDI - neurodevelopmental impairment
Table 1: Prenatal, Perinatal and Postnatal Characteristics of Study PopulationAbbreviations: IVH - intraventricular hemorrhage; NDI - neurodevelopmental impairment