Predicting Failure of Noninvasive Respiratory Support Using Deep Recurrent Learning

Background Noninvasive respiratory support is increasingly used to support patients with acute respiratory failure. However, noninvasive support failure may worsen outcomes compared to primary support with invasive mechanical ventilation. Therefore, there is a need to identify patients where noninvasive respiratory support is failing so that treatment can be reassessed and adjusted. The objective of this study was to develop and evaluate three recurrent neural network models to predict noninvasive respiratory support failure.

Methods This is a cross-sectional observational study to evaluate the ability of deep recurrent neural network models (long short-term memory, gated recurrent unit, and gated recurrent unit with trainable decay) to predict failure of noninvasive respiratory support. Data were extracted from electronic health records from all adult (≥ 18 years) patient records requiring any type of oxygen therapy or mechanical ventilation between November 1, 2013, and September 30, 2020 across 46 ICUs in the Southwest United States in a single healthcare network. Input variables for each model included serum chloride, creatinine, albumin, respiratory rate, heart rate, pulse oximetry oxygen saturation (SpO₂), fraction of inspired oxygen (FiO₂), oxygen saturation (SaO₂), and two measurements each (point-of-care and laboratory measurement) of partial pressure of carbon dioxide and partial pressure of arterial oxygen from an arterial blood gas.

Results Time series data from electronic health records were available for 22,075 patients. The highest accuracy and area under the receiver operating characteristic curve were for the long short-term memory model (94.04% and 0.9636, respectively). Accurate predictions were made 12 hours after ICU admission and performance remained high well in advance of noninvasive respiratory support failure.

Conclusion Recurrent neural network models using routinely collected time-series data can accurately predict noninvasive respiratory support failure well before intubation. This lead time may provide an opportunity to intervene to optimize patient outcomes.