Leveraging machine learning to study how temperament scores predict pre-term birth status

Abstract

Background: Preterm birth (birth at <37 completed weeks gestation) is a significant public heatlh concern worldwide. Important health, and developmental consequences of preterm birth include altered temperament development, with greater dysregulation and distress proneness.

Aims: The present study leveraged advanced quantitative techniques, namely machine learning approaches, to discern the contribution of narrowly defined and broadband temperament dimensions to birth status classification (full-term vs. preterm). Along with contributing to the literature addressing temperament of infants born preterm, the present study serves as a methodological demonstration of these innovative statistical techniques.

Study design: This study represents a metanalysis conducted with multiple samples (N = 19) including preterm (n = 201) children and (n = 402) born at term, with data combined across investigations to perform classification analyses.

Subjects: Participants included infants born preterm and term-born comparison children, either matched on chronological age or age adjusted for prematurity.

Outcome measures: Infant Behavior Questionnaire-Revised Very Short Form (IBQ-R VSF) was completed by mothers, with factor and item-level data considered herein.

Results and conclusions: Accuracy estimates were generally similar regardless of the comparison groups. Results indicated a slightly higher accuracy and efficiency for IBQR-VSF item-based models vs. factor-level models. Divergent patterns of feature importance (i.e., the extent to which a factor/item contributed to classification) were observed for the two comparison groups (chronological age vs. adjusted age) using factor-level scores; however, itemized models indicated that the two most critical items were associated with effortful control and negative emotionality regardless of comparison group.

Keywords: Infancy; Preterm birth; Quantitative methodology; Temperament.

Fig. 1.

a) R2 model comparisons for chronological age groups – itemized vs. factorized variables; b) area under the curve (AUC) model comparisons for chronological age groups – itemized vs. factorized variables; c) R2 model comparisons for age adjusted groups – itemized vs. factorized variables; d) area under the curve (AUC) model comparisons for age adjusted groups – itemized vs. factorized variables.

Fig. 2.

Random forest itemized variable feature importance, relative influence in classification, for: a) infant born pre-term vs. the chronological age control group; b) infant born preterm vs. the adjusted age control group.