Development of a Deep Learning Network to Classify Inferior Vena Cava Collapse to Predict Fluid Responsiveness.

Area of Science:

• Critical Care Medicine
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Fluid responsiveness assessment is crucial for managing critically ill patients.
• Inferior vena cava (IVC) collapsibility on ultrasound (US) is a potential indicator of fluid responsiveness.
• Predicting fluid responsiveness accurately can optimize patient outcomes.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm, specifically a long short-term memory (LSTM) network, for classifying IVC collapsibility from ultrasound videos.
• To assess the LSTM network's ability to predict fluid responsiveness in critically ill patients.
• To compare the LSTM network's performance against point-of-care US experts.

Main Methods:

• A dataset of IVC US videos from spontaneously breathing, critically ill patients undergoing fluid resuscitation was utilized.
• An LSTM network was trained on 90% of the video data and tested on the remaining 10%.
• Data augmentation techniques were employed to increase the effective training sample size.

Main Results:

• The LSTM network achieved an area under the receiver operating characteristic curve (AUC) of 0.70 for predicting fluid responsiveness.
• Positive and negative likelihood ratios were calculated to assess the predictive value of the model.
• Point-of-care US experts achieved a significantly higher AUC of 0.94 in comparison.

Conclusions:

• A deep learning LSTM network can be trained using IVC US videos to predict fluid responsiveness.
• The developed LSTM network demonstrated moderate performance, outperforming chance but underperforming expert clinicians.
• Further research with larger datasets is recommended to improve the LSTM network's accuracy and clinical utility.