Machine learning models did not achieve clinically acceptable performance (RMSE ≤ 5 cm H2O) for predicting expiratory occlusion pressure, with the best model yielding a test set RMSE of 7.9 cm H2O.
Observational (n=1,212)
Can machine learning models accurately predict expiratory occlusion pressure (Pocc) using bedside measurements and clinical data in mechanically ventilated patients?
Machine learning models based on standard bedside measurements and clinical data were unable to predict expiratory occlusion pressure with clinically acceptable accuracy in mechanically ventilated patients.
Abstract Rationale Both insufficient and excessive mechanical ventilation (MV) can predispose patients to lung and diaphragm injury due to perturbations in respiratory drive and effort. Respiratory effort is modulated by a complex interaction between ventilator settings, sedation, physiological characteristics, and illness characteristics. As a first step to developing clinical decision support for optimizing respiratory effort, we set out to test whether we could derive a machine learning (ML) model to predict the expiratory occlusion pressure (Pocc) based on bedside measurements and clinical data. Methods Using retrospective data from a longitudinal, observational study (6,308 patient-days in 1,212 patients), ML models were trained to predict Pocc based on MV settings, sedative agents, and clinical characteristics (sex, height, weight). The data were partitioned into an 80/20 train/test split for training and testing, respectively. Using the train dataset, we trained eleven different ML models and tuned their hyperparameters via Bayesian Optimization with 10-fold cross-validation (CV). To estimate the out-of-sample performance of the models, we performed nested CV with 5 inner folds and 5 outer folds. For external validation, we assessed the models’ performance on the test dataset. Model performance was assessed based on the root mean square error (RMSE) between predicted and observed Pocc values. Models that yielded an RMSE score ≤ 5 cm H2O met our criterion for clinically acceptable predictive performance. Results The five best-performing models and their respective nested cross-validation (NCV) and test set prediction (TS) RMSE scores in cm H2O were: Random forest (NCV = 7.4 cm H2O, TS = 7.9 cm H2O), Generalized linear regression with elastic net regularization (NCV = 7.2 cm H2O, TS = 8.1 cm H2O), Generalized additive model (NCV = 7.3 cm H2O, TS = 8.1 cm H2O), Extreme gradient boosting model (NCV = 7.3 cm H2O, TS = 8.1 cm H2O), and Bayesian additive regression tree (NCV = 8.1 cm H2O, TS = 8.2 cm H2O). The five most influential variables in the best-performing model (random forest) were: respiratory rate (total - set), tidal volume adjusted by predicted body weight, ventilator mode, dynamic driving pressure, and day of ICU admission. Conclusions None of the tested models met our criterion for clinically acceptable performance. Of the models that were assessed, the random forest model demonstrated the best performance. As might be expected on clinical grounds, respiratory rate, tidal volume, and mode were the strongest predictor variables in these models. This abstract is funded by: None
Iftikhar et al. (Fri,) conducted a observational in Mechanical ventilation (n=1,212). Machine learning models was evaluated on Root mean square error (RMSE) between predicted and observed Pocc values. Machine learning models did not achieve clinically acceptable performance (RMSE ≤ 5 cm H2O) for predicting expiratory occlusion pressure, with the best model yielding a test set RMSE of 7.9 cm H2O.