Developing and validating machine learning models to predict next-day extubation

Area of Science:

• Critical Care Medicine
• Biomedical Informatics
• Machine Learning in Healthcare

Background:

• Mechanical ventilation (MV) liberation criteria are often imprecise, leading to prolonged MV or reintubation.
• Adverse outcomes are associated with both prolonged MV and reintubation.
• Protocol-driven daily assessments expedite extubation but require dedicated staff.

Purpose of the Study:

• To determine if machine learning (ML) applied to electronic health records (EHR) can predict next-day extubation.
• To evaluate ML model performance in predicting extubation readiness.

Main Methods:

• Examined 37 clinical features from 12 AM-8 AM on ICU days from a prospective cohort.
• Utilized three data encoding/imputation strategies.
• Built and compared XGBoost, LightGBM, logistic regression, LSTM, and RNN models for prediction.
• Tested models on internal and external ICU cohorts.

Main Results:

• The best model (LSTM) achieved an AUROC of 0.870 in both internal and external test cohorts.
• Key predictors included plateau pressure and Richmond Agitation Sedation Scale (RASS) score.
• Models often predicted extubation readiness days before actual extubation (63.8% within 3 days).

Conclusions:

• ML models show promise as clinical decision support tools for mechanical ventilation liberation.
• ML models may assist in identifying patients ready for extubation earlier.
• Further randomized controlled trials are needed to confirm safety, efficacy, and cost-effectiveness compared to protocol-based care.