Smart alarms towards optimizing patient ventilation in intensive care: the driving pressure case

Alarms are a substantial part of clinical practice, warning clinicians of patient complications. In this paper, we focus on alarms in the intensive care unit and especially on the use of machine learning techniques for the creation of alarms for the ventilator support of patients. The aim is to study a method to enable timely interventions for intubated patients and prevent complications induced by high driving pressure (ΔP) and lung strain during mechanical ventilation.

The relation between the ΔP and the total set of the ventilator parameters was examined and resulted in a predictive model with bimodal implementation for the short-term prediction of the ΔP level (high/low). The proposed method includes two sub-models for the prediction of future ΔP level based on the current level being high or low, named cH and cL, respectively. Based on this method, for both sub-models, an alarm will be triggered when the predicted ΔP level is considered to be high. In this vein, three classifiers (the random forest, linear support vector machine, and kernel support vector machine methods) were tested for each sub-model. To adjust the highly unbalanced classes, four different sampling methods were considered: downsampling, upsampling, synthetic minority over-sampling technique (SMOTE) sampling, and random over-sampling examples (ROSE) sampling.

For the cL sub-model the combination of linear support vector machine with SMOTE sampling showed the best performance, resulting in accuracy of 93%, while the cH sub-model reached the best performance, with accuracy of 73%, with kernel support vector machine combined with the downsampling method.

The results are positive in terms of the generation of new alarms in mechanical ventilation. The technical and organizational possibility of integrating data from multiple modalities is expected to further advance this line of work.