Development and validation of a machine learning model for real-time prediction of invasive mechanical ventilation weaning readiness

Abstract

Purpose: To develop and validate a bedside machine learning (ML) decision support tool for prediction of invasive mechanical ventilation (IMV) weaning readiness.

Methods: Adults admitted after 2010 who underwent IMV (>24 h) were included from MIMIC-IV (development and internal validation) and AmsterdamUMCdb (external validation) databases. XGBoost boosted trees approach was used to develop three models predicting IMV weaning readiness within 24, 48, and 72 h by integrating electronic health record data. The areas under Receiver Operating Characteristic (auROC), the Precision-Recall curve (auPR) curves, and performance metrics were assessed. Sensitivity analyses evaluated the impact of gender, ethnicity, age and admission reason on model performance.

Results: 8565 patients from MIMIC-IV and 2626 from AmsterdamUMCdb were included. In the external validation cohort, the 24-, 48-, and 72-h models had auROCs of 0.847, 0.795 and 0.789, and auPR of 54.17, 54.56 and 59.4, respectively. Sensitivity was >0.75 for all models, but specificity decreased from 0.79 to 0.63 between the 24-h and 72-h models. Lower performances were observed for older (> 60 years) and neurosurgical patients.

Conclusions: This study presents three ML models for real-time prediction of IMV weaning readiness, offering a promising approach to enhance clinical decision-making and optimize patient care.

Keywords: Invasive mechanical ventilation; Machine-learning; Respiratory failure; Weaning.