Predicting Failure of Noninvasive Respiratory Support Using Deep Recurrent Learning.

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Respiratory Physiology

Background:

• Noninvasive respiratory support (NRS) is widely used for acute respiratory failure.
• Failure of NRS can lead to worse outcomes than primary invasive ventilation.
• Predicting NRS failure is crucial for timely treatment adjustments.

Purpose of the Study:

• To develop and evaluate recurrent neural network (RNN) models for predicting NRS failure.
• To assess the accuracy and lead time of RNN models in identifying patients at risk of NRS failure.

Main Methods:

• A cross-sectional observational study utilizing electronic health record data from 22,075 adult patients across 46 ICUs.
• Evaluation of deep RNN models, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU).
• Input variables included clinical and laboratory data such as serum electrolytes, creatinine, albumin, vital signs, and blood gas parameters.

Main Results:

• The LSTM model achieved the highest accuracy (94.04%) and area under the receiver operating characteristic curve (0.9636).
• Accurate predictions of NRS failure were made up to 12 hours after ICU admission.
• Model performance remained high, providing significant lead time before actual failure.

Conclusions:

• RNN models can accurately predict noninvasive respiratory support failure using routinely collected time-series data.
• The predictive lead time allows for early intervention and optimization of patient care.
• This approach may improve outcomes for patients with acute respiratory failure.