PP444 Topic: AS15–Lung: Respiratory Support/Acute Respiratory Failure/Other: TOWARDS AI FOR EXTUBATION FAILURE PREDICTION: A RETROSPECTIVE COHORT STUDY FROM TWO UK PAEDIATRIC INTENSIVE CARE UNITS

Le résumé fourni par la source

Aims & Objectives: 5-15% of children extubated after mechanical ventilation require re-intubation within 48 hours. Extubation failure is associated with increased morbidity, mortality, and longer ICU length of stay. Thus, predicting children with higher risk of extubation failure is desirable. Methods: Retrospective cohort study using data from electronic medical records (Philips) to identify the importance of key variables (demographics, lab results, vital signs, ventilator settings) as predictors of extubation failure. Patients aged 0-18 yrs admitted to PICU at St Mary’s Hospital (general PICU) or the Royal Brompton Hospital (cardio-respiratory PICU) in London 2013-2022, receiving invasive mechanical ventilation via endotracheal tube for >6 hours, were eligible. Model methodology involved recursive feature elimination, over-resampling and cross-validation. Ethical approval was obtained via IRAS. Results: 4,488 patients were identified, with 3.3% PICU mortality. Median length of mechanical ventilation was 121 hours (IQR 64-229); the calculated rate of extubation failure was 5.5%. There were small but significant differences in physiological and ventilators settings (e.g. PIP, deltaP, RR) immediately pre-extubation between those who were successfully extubated and those who required re-intubation. A logistic regression model incorporating all significant variables achieved an area under the receiver-operating-characteristic-curve (AUROC) of 0.65. Machine learning approaches (Figure 1) ranged from AUROC 0.55-0.69. Conclusions: This study returned an extubation failure rate of 5.5% within 48 hours. Models incorporating significant predictor variables only had moderate predictive ability. This study illustrates the challenge of using routinely collected data to model which patients are likely to fail extubation, despite inclusion of of physiological variables and ventilation settings. Keywords: modelling, Machine Learning, extubation, intubation

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.