PREDICTION OF PHYSIOLOGICAL DETERIORATION AND MORTALITY IN MECHANICALLY VENTILATED PATIENTS ADMITTED TO THE ICU

Abstract

TOPIC: Critical Care TYPE: Original Investigations PURPOSE: Recent studies have shown that the severity of lung damage in experimental models of ventilator-induced lung injury depend on the mechanical power (MP) of the ventilator, which describes the energy delivered by the ventilator to the respiratory system per unit time. This study aims to build machine learning models that leverage knowledge of ICU data and ventilator settings to predict physiological deterioration and mortality in mechanically ventilated patients. METHODS: Patient data from the Phillips eICU Database were used to build predictive models. Inclusion criteria include age ≥18 years, ICU stay duration ≥48 hours, ventilation duration ≥48 hours, and volume- or pressure-controlled mechanical ventilation. Physiological deterioration was defined as any increase between daily Sequential Organ Failure Assessment (SOFA) scores within the first 7 days of ventilation. Daily maximum MP values were calculated using previously validated equations. Predictive features such as ventilator settings, lab values, medications, vitals, comorbidity indices, patient demographics, and hospital stay information were extracted for each patient stay. Classification models including logistic regression, random forest, and support-vector machines were evaluated for predicting mortality and physiological deterioration using area under the receiver operating characteristic curve (AUC) with six-fold cross-validation. RESULTS: A total of 1338 patients met inclusion criteria for this study. After data processing and imputation, features from 789 patients were curated as inputs to predictive models. These preliminary models showed predictive power for overall physiological deterioration with an AUC of 0.74 ± 0.03. For organ-specific deterioration, these models exhibited comparable or higher performance in predicting renal deterioration (AUC: 0.74 ± 0.08) and pulmonary deterioration (AUC: 0.79 ± 0.03; p<0.05). Prediction of in-hospital mortality resulted in an AUC of 0.78 ± 0.05. CONCLUSIONS: Based on our preliminary data, these predictive models have the potential to be used in real-time to enable early detection and intervention for patients at high risk of physiological deterioration or in-hospital mortality. Implementation of additional time-series physiological data is expected to further improve the performance of these models. CLINICAL IMPLICATIONS: This study explores the effectiveness of multiple statistical learning models in predicting physiological deterioration and mortality in mechanically ventilated ICU patients. Our results have the potential to inform clinicians about the risk of downstream complications associated with a patient's ICU stay and can therefore aid in the prevention of end-organ dysfunction associated with mechanical ventilation. With models that have predictive capacity days before the onset of physiological deterioration, clinicians and other health care providers in the ICU will be able to provide more personalized care for mechanically ventilated patients. We foresee that these predictive models will not only minimize the risk of or prevent mechanical ventilation-associated adverse outcomes but will also be useful in risk-stratifying mechanically ventilated patients based on patient-specific and ventilator-derived data. DISCLOSURES: No relevant relationships by Timothy Bedard, source=Web Response No relevant relationships by Yunru Chen, source=Web Response No relevant relationships by Andy Ding, source=Web Response No relevant relationships by Joseph Greenstein, source=Web Response No relevant relationships by Madi Kusmanov, source=Web Response No relevant relationships by Pedro Mendez-Tellez, source=Web Response No relevant relationships by Morgan Sanchez, source=Web Response No relevant relationships by Shreyash Sonthalia, source=Web Response No relevant relationships by Raimond Winslow, source=Web Response