This retrospective study aimed to explore the effects of personalized mechanical power (MP) thresholds on ICU mortality among mechanically ventilated patients. The study also aimed to devise an individualized approach for adjusting mechanical ventilation settings based on MP and other covariables. Data were sourced from the Amsterdam University Medical Centers Database (AmsterdamUMCdb), a repository resulting from the collaborative effort between the Society of Critical Care Medicine and the European Society of Intensive Care Medicine(10).
1. Study population:
The study encompassed patients aged 18 years or older who needed a minimum of 48 hours of pressure-controlled mechanical ventilation. Exclusions encompassed patients who died or were extubated within the initial 48 hours, as well as those with inadequate data. Based on their PaO2/FiO2 ratio (PFR) over the first six hours of mechanical ventilation, patients were stratified into distinct hypoxemia categories – nonhypoxemic (PFR > 300), mildly hypoxemic (PFR 200-300), moderately hypoxemic (PFR 100-199), and severely hypoxemic (PFR < 100). The study population was analyzed with regard to mortality outcomes.
2. Data Collection and Mechanical Power Calculation:
Baseline patient characteristics were meticulously gathered, while mechanical ventilation parameters were also recorded. The time-weighted average (TWA) was calculated over the initial 24 and 48 hours of mechanical ventilation. The mechanical power of ventilation was determined using the surrogate formula(11):
The TWA of MP over the first 24 and 48 hours of mechanical ventilation was then normalized to the patient's ideal body weight (IBW). This metric provides insight into the potential lung tissue damage associated with mechanical ventilation.
3. Statistical analysis and machine learning models:
The collected data were described using the mean, standard deviation (SD), median, and interquartile range (IQR) for symmetric and nonsymmetric distributions, respectively. The Mann‒Whitney test was employed for nonsymmetric continuous variables. In contrast, the T-tailed test was used for symmetrically distributed variables to compare patients grouped by mortality status, and chi-square was used for categorical data. A P value < 0.05 indicates statistical significance. Unadjusted odds ratios (ORs) were calculated to assess the influence of TWA mechanical power normalized to IBW on mortality.
Safe upper limits for IBW-adjusted MP were determined for each hypoxemia group by applying the Mann-Whitney test. It was applied to compare the distributions of IBW-adjusted MP between mortality groups, focusing on the alternative hypothesis that the variable's distribution was greater among nonsurvivors. Employing a predefined significance level, the test results were evaluated. If the p value was deemed significant, a substantial upper limit value was calculated using the survivors' percentile corresponding to the significance level.
Survival (COX) analysis comparing populations grouped according to the identified limits was performed, including noncollinear variables that had a significant association with mortality. Kaplan‒Meier estimator plotted for visualization.
3.1. Mortality Prediction:
First, imputation was performed for data with 10% missing values or less via k-nearest neighbour (k=7). Afterward, the synthetic minority oversampling technique was used to balance the data between nonsurvivors and survivors(12). Ventilation-related features are included within the model by default, i.e., MP, IBW-normalized MP, IBW-normalized tidal volume, 48-h standard deviation of MP, PEEP, respiratory rate, and driving pressure. In addition, relevant features are selected via 2 steps: 1) removing features that demonstrated larger than 90% cross-correlation and 2) selecting the 10 most significant features based on random forest. Ultimately, various machine learning models were systematically evaluated for their ability to predict ICU mortality, including logistic regression, random forest, SVM, Ada boosting, xgBoost, and stacking(13–16). Performance evaluation utilized accuracy, precision, recall, and the area under the receiver operating curve (ROC-AUC) to gauge the predictive capabilities of each model.
4. Individualization of MV Settings:
The individualization of the MV setting in the present study aims to trade off over- and underventilation by leveraging the mortality prediction model from the previous section. Figure 1 illustrates the optimization scheme. A patient is admitted to the ICU, and the responsible physician manually sets the MV settings. The respiratory acidosis status of the patient is then observed, e.g., by taking measures of end-tidal CO2 or PaCO2. The MV settings are then optimized based on whether the degree of potential acidosis is high or low to maximize minute ventilation (VE) or minimize MP stepwise, respectively. This step provides optimized values for the tidal volume and/or driving pressure and respiratory rate. The stepwise minimization of MP ultimately aims to wean the patient systematically. In a parallel workflow, PEEP titration is recommended to find the optimal PEEP setting for each patient(17), thereby significantly reducing the degree of freedom within the optimization process. The recommended MV settings and other measured covariables are then fed into the mortality prediction. The whole process of MV setting optimization is performed iteratively until the mortality prediction results in false.