Extubation Readiness in Preterm Infants

Clinical Decision Support from Time Series Signal Data

Machine learning analysis has the potential to enhance the prediction accuracy of extubation readiness in preterm infants while utilizing readily available data streams from bedside pulse oximeters and ventilators.

The objective of this study was to predict extubation readiness in preterm infants using machine learning analysis of bedside pulse oximeter and ventilator data. This observational study used prospective recordings of oxygen saturation (SpO2) and ventilator data from infants <30 weeks of gestation age. Research pulse oximeters collected SpO2 (1 Hz sampling rate) to quantify intermittent hypoxemia (IH). Continuous ventilator metrics were collected from bedside ventilators. Data modeling was completed using unbiased machine learning algorithms. Three model sets were created using the following data source combinations: (1) IH and ventilator (IH + SIMV), (2) IH, and (3) ventilator (SIMV). Infants were also analyzed separated by postnatal age (infants <2 or ≥2 weeks of age). Models were compared by area under the receiver operating characteristic curve (AUC).

Machine learning analysis has the potential to enhance the prediction accuracy of extubation readiness in preterm infants while utilizing readily available data streams from bedside pulse oximeters and ventilators.

Three model sets were developed using different data streams. The first model used a combination of pulse oximeter and ventilator data (IH + SIMV). The other two models included pulse oximeter only (IH) or ventilator (SIMV) only data. Demographic birth information and data from time of extubation (e.g. age, weight, sex, race) were included in all analyses. The study was approved by the institutional review board of the University of Kentucky. A waiver of informed consent was obtained to utilize machine learning techniques for evaluating extubation readiness in preterm infants.

A total of 110 extubation events from 110 preterm infants were analyzed. Infants had a median gestation age and birth weight of 26 weeks and 825 g, respectively. Of the 3 models presented, the IH + SIMV model achieved the highest AUC of 0.77 for all infants. Separating infants by postnatal age increased accuracy further achieving AUC of 0.94 for <2 weeks of age group and AUC of 0.83 for ≥2 weeks group.

The variables from both the pulse oximeter and ventilator were input based on computed percentiles calculated from the data set. Percentiles were separated into quantiles defined at 10%, 25%, 50% (median), 75%, and 90%. Missing data did not contribute to percentiles. The counts were performed by counting the number of rows in the data for a given variable for each patient. We also computed fractions by dividing the counts by the number of rows during the interval. The percentiles, fractions, and counts were computed on a time window prior to the extubation event. Windows of 2 hours, 4 hours, 8 hours, 16 hours, and 24 hours were evaluated.

The following candidate machine learning algorithms were evaluated using scikit-learn as noted:

1. Random Forest (scikit-learn RandomForestClassifier)
2. Neural Network (scikit-learn MLPClassifier)
3. XGBoost
4. Bagging (BalancedBaggingClassifier)
5. Linear SVM (scikit-learn SGDClassifier)
6. Logistic Regression (scikit-learn LogisticRegression)

Models were compared using area under the ROC curve (AUC), sensitivity, and specificity on a 3-fold cross-validation. Hyperparameter tuning was not performed due to the small number of positive samples in some subcategories.

Shapley Additive exPlanations (SHAP) analysis reviewed for IH + SIMV, IH, and SIMV models of the population of all infants as well as the subgroups of infants < 2 weeks and ≥ 2 weeks of age. SHAP analyses assign weight to the top performing features for a model’s prediction.

Of the machine learning algorithms used, random forest had the most consistent performance as indicated by the highest average AUC throughout data sources and population analyses.

The data used for this project can be found in Predicting Extubation Readiness in Preterm Infants Utilizing Machine Learning: A Diagnostic Utility Study, a paper published in the Journal of Pediatrics.

The Journal of Pediatrics published Predicting Extubation Readiness in Preterm Infants Utilizing Machine Learning: A Diagnostic Utility Study.

This study is supported in part by the National Center for Advancing Translational Sciences (UL1TR001998 to E.G.A.J), NIH NICHD (K23HD109471 to E.G.A.J), and the University of Kentucky College of Medicine Dean’s Office. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or the University of Kentucky.

Dr. Cody Bumgardner and Samuel Armstrong worked on Extubation Readiness in Preterm Infants.

We thank the support of the University of Kentucky NICU nurses and research staff.