AI Driven Prediction Model for Mechanical Ventilation Duration Using Lung Ultrasound Score

Abstract

Background: Extended mechanical ventilation drives up healthcare costs and simultaneously increases the risk of complications, adding to the patient’s overall health burden. Accurate prediction of mechanical ventilation duration by clinicians remains difficult in present-day critical care practice. Accurately predicting this duration can support timely clinical decisions, such as resource allocation and early tracheostomy planning; where lung ultrasound studies may show promise. Objective: To develop a machine learning (ML) model for predicting mechanical ventilation duration with Lung ultrasound score. Methodology: An ML based model was developed and tested using 218 mechanically ventilated patients from ICUs at AIIMS, Delhi. The classification model was developed to predict short-($\boldsymbol{\leq} \mathbf{3}$ days) or long-term ($\boldsymbol{\gt} \mathbf{3}$ days) ventilation requirements using the feature-Lung ultrasound score. Models were trained using 5-fold cross-validation with 8020 training and test division. The model was evaluated using accuracy and Area under the Receiver Operating Characteristic Curve (AUROC). Results: The model achieved training accuracy of $\mathbf{6 0. 0 - 6 3. 4 \%}$ (ROC: $\mathbf{0. 5 9 - 0. 6 4}$) and test accuracy of 72.1-74.4% (ROC: 0.72-0.81). Conclusion: The automated prediction of mechanical ventilation duration aids early identification of prolonged ventilation, optimizing ICU workflows, guiding personalized care, and supporting clinical decisions in resource-limited settings.

Ähnliche Arbeiten