BACKGROUND: Ventilator weaning protocols are commonly implemented for patients receiving mechanical ventilation. However, despite such protocols, the rate of extubation failure remains high. This study analyzed the usefulness and accuracy of machine learning in predicting extubation success. METHODS: We retrospectively evaluated data from patients who underwent intubation for respiratory failure and received mechanical ventilation in an intensive care unit (ICU). Information on 57 features, including patient demographics, vital signs, laboratory data, and ventilator data, were extracted. Extubation failure was defined as re-intubation within 72 hours of extubation. For supervised learning, data were labeled as intubation-required or not. We used three learning algorithms (Random Forest, XGBoost, and LightGBM) to predict successful extubation. We also analyzed important features and evaluated the area under curve (AUC) and prediction metrics. RESULTS: Overall, 13 of the 117 included patients required re-intubation. LightGBM had the highest AUC (0.950), followed by XGBoost (0.946) and Random Forest (0.930). The accuracy, precision, and recall performance were 0.897, 0.910, and 0.909 for Random Forest; 0.910, 0.912, and 0.931 for XGBoost; and 0.927, 0.915, and 0.960 for LightGBM, respectively. The most important feature was duration of mechanical ventilation, followed by fraction of inspired oxygen, positive end-expiratory pressure, maximum and mean airway pressures, and Glasgow Coma Scale. CONCLUSIONS: Machine learning predicted successful extubation of ICU patients on mechanical ventilation. LightGBM had the best overall performance. Duration of mechanical ventilation was the most important feature in all models.
BACKGROUND: Ventilator weaning protocols are commonly implemented for patients receiving mechanical ventilation. However, despite such protocols, the rate of extubation failure remains high. This study analyzed the usefulness and accuracy of machine learning in predicting extubation success. METHODS: We retrospectively evaluated data from patients who underwent intubation for respiratory failure and received mechanical ventilation in an intensive care unit (ICU). Information on 57 features, including patient demographics, vital signs, laboratory data, and ventilator data, were extracted. Extubation failure was defined as re-intubation within 72 hours of extubation. For supervised learning, data were labeled as intubation-required or not. We used three learning algorithms (Random Forest, XGBoost, and LightGBM) to predict successful extubation. We also analyzed important features and evaluated the area under curve (AUC) and prediction metrics. RESULTS: Overall, 13 of the 117 included patients required re-intubation. LightGBM had the highest AUC (0.950), followed by XGBoost (0.946) and Random Forest (0.930). The accuracy, precision, and recall performance were 0.897, 0.910, and 0.909 for Random Forest; 0.910, 0.912, and 0.931 for XGBoost; and 0.927, 0.915, and 0.960 for LightGBM, respectively. The most important feature was duration of mechanical ventilation, followed by fraction of inspired oxygen, positive end-expiratory pressure, maximum and mean airway pressures, and Glasgow Coma Scale. CONCLUSIONS: Machine learning predicted successful extubation of ICU patients on mechanical ventilation. LightGBM had the best overall performance. Duration of mechanical ventilation was the most important feature in all models.
Entities:
Keywords:
extubation failure; intensive care unit; machine learning; mechanical ventilation
Authors: Lucas M Fleuren; Tariq A Dam; Michele Tonutti; Daan P de Bruin; Robbert C A Lalisang; Diederik Gommers; Olaf L Cremer; Rob J Bosman; Sander Rigter; Evert-Jan Wils; Tim Frenzel; Dave A Dongelmans; Remko de Jong; Marco Peters; Marlijn J A Kamps; Dharmanand Ramnarain; Ralph Nowitzky; Fleur G C A Nooteboom; Wouter de Ruijter; Louise C Urlings-Strop; Ellen G M Smit; D Jannet Mehagnoul-Schipper; Tom Dormans; Cornelis P C de Jager; Stefaan H A Hendriks; Sefanja Achterberg; Evelien Oostdijk; Auke C Reidinga; Barbara Festen-Spanjer; Gert B Brunnekreef; Alexander D Cornet; Walter van den Tempel; Age D Boelens; Peter Koetsier; Judith Lens; Harald J Faber; A Karakus; Robert Entjes; Paul de Jong; Thijs C D Rettig; Sesmu Arbous; Sebastiaan J J Vonk; Mattia Fornasa; Tomas Machado; Taco Houwert; Hidde Hovenkamp; Roberto Noorduijn Londono; Davide Quintarelli; Martijn G Scholtemeijer; Aletta A de Beer; Giovanni Cinà; Adam Kantorik; Tom de Ruijter; Willem E Herter; Martijn Beudel; Armand R J Girbes; Mark Hoogendoorn; Patrick J Thoral; Paul W G Elbers Journal: Crit Care Date: 2021-12-27 Impact factor: 9.097