Predicting ICU Mortality in Acute Respiratory Distress Syndrome Patients Using Machine Learning: The Predicting Outcome and STratifiCation of severity in ARDS (POSTCARDS) Study.

OBJECTIVES: To assess the value of machine learning approaches in the development of a multivariable model for early prediction of ICU death in patients with acute respiratory distress syndrome (ARDS). DESIGN: A development, testing, and external validation study using clinical data from four prospective, multicenter, observational cohorts. SETTING: A network of multidisciplinary ICUs. PATIENTS: A total of 1,303 patients with moderate-to-severe ARDS managed with lung-protective ventilation. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: We developed and tested prediction models in 1,000 ARDS patients. We performed logistic regression analysis following variable selection by a genetic algorithm, random forest and extreme gradient boosting machine learning techniques. Potential predictors included demographics, comorbidities, ventilatory and oxygenation descriptors, and extrapulmonary organ failures. Risk modeling identified some major prognostic factors for ICU mortality, including age, cancer, immunosuppression, Pa o2 /F io2 , inspiratory plateau pressure, and number of extrapulmonary organ failures. Together, these characteristics contained most of the prognostic information in the first 24 hours to predict ICU mortality. Performance with machine learning methods was similar to logistic regression (area under the receiver operating characteristic curve [AUC], 0.87; 95% CI, 0.82-0.91). External validation in an independent cohort of 303 ARDS patients confirmed that the performance of the model was similar to a logistic regression model (AUC, 0.91; 95% CI, 0.87-0.94). CONCLUSIONS: Both machine learning and traditional methods lead to promising models to predict ICU death in moderate/severe ARDS patients. More research is needed to identify markers for severity beyond clinical determinants, such as demographics, comorbidities, lung mechanics, oxygenation, and extrapulmonary organ failure to guide patient management.

A Quantile Analysis of Plateau and Driving Pressures: Effects on Mortality in Patients With Acute Respiratory Distress Syndrome Receiving Lung-Protective Ventilation.

OBJECTIVES: The driving pressure (plateau pressure minus positive end-expiratory pressure) has been suggested as the major determinant for the beneficial effects of lung-protective ventilation. We tested whether driving pressure was superior to the variables that define it in predicting outcome in patients with acute respiratory distress syndrome. DESIGN: A secondary analysis of existing data from previously reported observational studies. SETTING: A network of ICUs. PATIENTS: We studied 778 patients with moderate to severe acute respiratory distress syndrome. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: We assessed the risk of hospital death based on quantiles of tidal volume, positive end-expiratory pressure, plateau pressure, and driving pressure evaluated at 24 hours after acute respiratory distress syndrome diagnosis while ventilated with standardized lung-protective ventilation. We derived our model using individual data from 478 acute respiratory distress syndrome patients and assessed its replicability in a separate cohort of 300 acute respiratory distress syndrome patients. Tidal volume and positive end-expiratory pressure had no impact on mortality. We identified a plateau pressure cut-off value of 29 cm H2O, above which an ordinal increment was accompanied by an increment of risk of death. We identified a driving pressure cut-off value of 19 cm H2O where an ordinal increment was accompanied by an increment of risk of death. When we cross tabulated patients with plateau pressure less than 30 and plateau pressure greater than or equal to 30 with those with driving pressure less than 19 and driving pressure greater than or equal to 19, plateau pressure provided a slightly better prediction of outcome than driving pressure in both the derivation and validation cohorts (p < 0.0000001). CONCLUSIONS: Plateau pressure was slightly better than driving pressure in predicting hospital death in patients managed with lung-protective ventilation evaluated on standardized ventilator settings 24 hours after acute respiratory distress syndrome onset.

A clinical classification of the acute respiratory distress syndrome for predicting outcome and guiding medical therapy*.

OBJECTIVE: Current in-hospital mortality of the acute respiratory distress syndrome (ARDS) is above 40%. ARDS outcome depends on the lung injury severity within the first 24 hours of ARDS onset. We investigated whether two widely accepted cutoff values of PaO2/FIO2 and positive end-expiratory pressure (PEEP) would identify subsets of patients with ARDS for predicting outcome and guiding therapy. DESIGN: A 16-month (September 2008 to January 2010) prospective, multicenter, observational study. SETTING: Seventeen multidisciplinary ICUs in Spain. PATIENTS: We studied 300 consecutive, mechanically ventilated patients meeting American-European Consensus Conference criteria for ARDS (PaO2/FIO2 ≤ 200 mm Hg) on PEEP greater than or equal to 5 cm H2O, and followed up until hospital discharge. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Based on threshold values for PaO2/FIO2 (150 mm Hg) and PEEP (10 cm H2O) at ARDS onset and at 24 hours, we assigned patients to four categories: group I (PaO2/FIO2 ≥ 150 on PEEP < 10), group II (PaO2/FIO2 ≥ 150 on PEEP ≥ 10), group III (PaO2/FIO2 < 150 on PEEP < 10), and group IV (PaO2/FIO2 < 150 on PEEP ≥ 10). The primary outcome was all-cause in-hospital mortality. Overall hospital mortality was 46.3%. Although at study entry, patients with PaO2/FIO2 less than 150 had a higher mortality than patients with a PaO2/FIO2 greater than or equal to 150 (p = 0.044), there was minimal variability in mortality among the four groups (p = 0.186). However, classification of patients in each group changed markedly after 24 hours of usual care. Group categorization at 24 hours provided a strong association with in-hospital mortality (p < 0.00001): group I had the lowest mortality (23.1%), whereas group IV had the highest mortality (60.3%). CONCLUSIONS: The degree of lung dysfunction established by a PaO2/FIO2 of 150 mm Hg and a PEEP of 10 cm H2O demonstrated that ARDS is not a homogeneous disorder. Rather, it is a series of four subsets that should be considered for enrollment in clinical trials and for guiding therapy. A major contribution of our study is the distinction between survival after 24 hours of care versus survival at the time of ARDS onset.

Respiratory Subsets in Patients with Moderate to Severe Acute Respiratory Distress Syndrome for Early Prediction of Death.

Introduction: In patients with acute respiratory distress syndrome (ARDS), the PaO2/FiO2 ratio at the time of ARDS diagnosis is weakly associated with mortality. We hypothesized that setting a PaO2/FiO2 threshold in 150 mm Hg at 24 h from moderate/severe ARDS diagnosis would improve predictions of death in the intensive care unit (ICU). Methods: We conducted an ancillary study in 1303 patients with moderate to severe ARDS managed with lung-protective ventilation enrolled consecutively in four prospective multicenter cohorts in a network of ICUs. The first three cohorts were pooled (n = 1000) as a testing cohort; the fourth cohort (n = 303) served as a confirmatory cohort. Based on the thresholds for PaO2/FiO2 (150 mm Hg) and positive end-expiratory pressure (PEEP) (10 cm H2O), the patients were classified into four possible subsets at baseline and at 24 h using a standardized PEEP-FiO2 approach: (I) PaO2/FiO2 ≥ 150 at PEEP < 10, (II) PaO2/FiO2 ≥ 150 at PEEP ≥ 10, (III) PaO2/FiO2 < 150 at PEEP < 10, and (IV) PaO2/FiO2 < 150 at PEEP ≥ 10. Primary outcome was death in the ICU. Results: ICU mortalities were similar in the testing and confirmatory cohorts (375/1000, 37.5% vs. 112/303, 37.0%, respectively). At baseline, most patients from the testing cohort (n = 792/1000, 79.2%) had a PaO2/FiO2 < 150, with similar mortality among the four subsets (p = 0.23). When assessed at 24 h, ICU mortality increased with an advance in the subset: 17.9%, 22.8%, 40.0%, and 49.3% (p < 0.0001). The findings were replicated in the confirmatory cohort (p < 0.0001). However, independent of the PEEP levels, patients with PaO2/FiO2 < 150 at 24 h followed a distinct 30-day ICU survival compared with patients with PaO2/FiO2 ≥ 150 (hazard ratio 2.8, 95% CI 2.2−3.5, p < 0.0001). Conclusions: Subsets based on PaO2/FiO2 thresholds of 150 mm Hg assessed after 24 h of moderate/severe ARDS diagnosis are clinically relevant for establishing prognosis, and are helpful for selecting adjunctive therapies for hypoxemia and for enrolling patients into therapeutic trials.