Abstract
Preterm birth is the leading cause of mortality in children under the age of five. In particular, low birth weight and low gestational age are associated with an increased risk of mortality. Preterm birth also increases the risks of several complications, which can increase the risk of death, or cause long-term morbidities with both individual and societal impacts. In this work, we use machine learning for prediction of neonatal mortality as well as neonatal morbidities of bronchopulmonary dysplasia (BPD), necrotizing enterocolitis (NEC), and retinopathy of prematurity (ROP), among very low birth weight (VLBW) infants. Our predictors include time series data and clinical variables collected at the neonatal intensive care unit (NICU) of Children’s Hospital, Helsinki University Hospital. We examine 9 different classifiers and present our main results in AUROC and F1-score, which we propose for classifier selection in this study. We also investigate how the predictive performance of the classifiers evolves as the length of time series is increased, and examine the relative importance of different features using the random forest classifier, which we found to generally perform the best in all tasks. Our systematic study also involves different data preprocessing methods which can be used to improve classifier sensitivities. Our best classifier AUROC is 0.922 in mortality, 0.899 in BPD, 0.806 in NEC, and 0.846 in ROP.
Introduction
Preterm birth remains a critical global health challenge, particularly for infants born with very low birth weight (VLBW, <1500g). These infants face high risks of mortality and severe morbidities, including chronic lung disease (BPD), intestinal inflammation (NEC), and potential blindness (ROP). Traditional clinical practice relies on scoring systems like SNAP-II, which utilize a narrow 12-hour data window and simple logistic regression. Such models often fail to capture the dynamic physiological changes in neonates.
This study explores how machine learning can leverage high-frequency NICU data to provide more sensitive, data-driven neonatal care. A significant challenge in this domain is class imbalance; adverse outcomes are relatively rare, meaning standard metrics like AUROC can be deceptively high while the model fails to identify at-risk infants. We address this by proposing the F1-score as a primary metric for model selection and investigating the impact of time series length on predictive accuracy.
Methodology
Data and Preprocessing
The study utilized a pseudonymized dataset of 977 VLBW infants from the Helsinki University Hospital NICU (1999–2013). Features included static clinical variables (birth weight, gestational age) and physiological time series (heart rate, oxygen saturation, and blood pressure). Irregular sensor measurements were transformed into a synthetic synchronous dataset using 2-minute intervals and forward-filling.
Classification Models
We evaluated nine classification algorithms across four categories:
- Linear/Quadratic: Logistic Regression (LR), LDA, and QDA.
- Instance-based/Kernel: k-Nearest Neighbors (KNN) and Support Vector Machines (SVM).
- Probabilistic: Gaussian Processes (GP) with RBF, Matérn 3/2, and Matérn 5/2 kernels.
- Ensemble: Random Forest (RF).
Results and Discussion
The Random Forest (RF) classifier emerged as the most robust performer across nearly all tasks. A key finding was the importance of the F1-score for model selection. As shown in Figure 1, while AUROC values remained consistently high across many models, F1-scores varied significantly. By optimizing for F1-score and utilizing class sub-sampling (Experiment 2 in the figure), we achieved substantial improvements in sensitivity, ensuring fewer high-risk cases were missed.
Predictive Performance
- Mortality: RF achieved an AUROC of 0.922 and an F1-score of 0.493.
- BPD: RF achieved an AUROC of 0.899 and an F1-score of 0.704.
- NEC: This proved the most difficult to predict (AUROC 0.806, F1 0.215), likely due to its slower development. Performance improved as the observation window increased to 72 hours.
- ROP: RF achieved an AUROC of 0.846 and an F1-score of 0.368.
Feature Importance and Temporal Dynamics
For mortality and BPD, 12 hours of data was sufficient for high accuracy. However, physiological variability proved to be a critical signal for specific morbidities. For ROP, the standard deviation of diastolic blood pressure was a stronger predictor than static metrics like birth weight. Additionally, we found that the first 6 hours of life—often considered a noisy "adaptation period"—contain valuable signals that should not be excluded from modeling.
Conclusion
This systematic study demonstrates that machine learning, particularly Random Forest models optimized for F1-score, can significantly enhance the prediction of neonatal outcomes. While birth weight and gestational age remain vital, the inclusion of time series variability offers new avenues for early detection. Future work will focus on validating these models across multiple centers and moving toward real-time monitoring applications in the NICU.
Summary updated