564 - A pilot study evaluating the sole use of continuous ventilator and vital sign data to predict extubation success in neonates

Poster Number: 564
Publication Number: 564.2947

Background: Extubation failure in neonates is associated with increased risk of death and prolonged mechanical ventilation. Previous logistic regression and machine learning (ML) prediction models showed limited success.

Objective: Our objective was to evaluate whether incorporating continuous ventilator and vital sign data into a ML extubation prediction model improves model performance.

Design/Methods: This was a retrospective analysis of continuous vital sign and ventilator data in patients admitted since March 2020 to the NICU with ≥1 hour of data prior to an extubation attempt. Compassionate and extubations directly from high frequency ventilation were excluded. Extubation failure was defined as the need for reintubation within 14 days.

A hybrid ML approach was developed to analyze both temporal (vital sign and ventilator data) and non-temporal (clinical features and 24 hours of laboratory and medication) data. A gated recurrent unit based recurrent neural network (GRU-RNN) was trained to analyze up to 6 hours of time-series data combined with clinical features to predict extubation success or failure. Predictive performances of variations of this model were compared to conventional logistic regression and random forest models that used descriptive characteristics (mean, max, min, etc) instead of an ML approach to temporal data.

Results: 283 patients met inclusion criteria with 20 (7%) meeting the definition of extubation failure. The median time for extubation failure was 26.25 hours with only 2 patients requiring respiratory reintubation after 10 days of extubation. Table 1 shows higher prevalence of high-risk clinical features such as lower birthweight and gestational age (GA) in the extubation failure cohort. Table 2 shows the performance characteristics of the different models; all models performed well at predicting extubation success (precision 0.94—0.96, F1 score 0.89—0.95), but none predicted extubation failure accurately. The GRU-RNN models excluding GA and clinical parameters performed similar to the other models.

Conclusion(s): Due to unbalanced data, none of the models were able to predict extubation failure accurately while performance of extubation success was excellent. Performance of the model using only continuous ventilator and vital sign data (no clinical parameters) was comparable to models with high-risk clinical parameters included, suggesting that continuous data may contribute significantly to successful extubation prediction. These models may be useful in extubating infants sooner. Future multi-center work with a larger sample size of patients is needed to validate these findings.