Improving Mechanical Ventilator Clinical Decision Support Systems with A Machine Learning Classifier for Determining Ventilator Mode

Introduction

Mechanical ventilation (MV) is a life-saving intervention delivered in the intensive care unit (ICU) to patients with acute respiratory failure. When delivered properly, MV can allow injured lungs heal while the ventilator performs the majority of work of breathing for a patient.

When delivered improperly, MV has been associated with a variety of adverse clinical outcomes including patient discomfort, increased sedative dosing, longer ICU length of stay, increased chance of ventilatorinduced lung injury, and lower survival , .

One such piece of information that many MV CDSS lack is the choice of ventilation mode (VM) that determines the pattern of flow and pressure delivery with each breath (Figure 1 B-D ).

Methodology

To the best of our knowledge, only one previous effort has developed a rulebased classifier using analysis of VWD for providing hourly VM classifications. Many patients had 2-4 hour periods selected where ventilator mode was switched multiple times, Other modes such as pressure regulated volume control (PRVC), volume support, and airway pressure release ventilation (APRV) were found, and annotated within these epochs, but were excluded in our final analysis because.

Study Design

Using both Scikitlearn and Pytorch ML libraries , , we then evaluated the use of multiple ML algorithms including: support vector machine (SVM) , multi-layer perceptron (MLP), long-short term memory recurrent neural network (LSTM RNN) , logistic regression, and a random forest (RF) classifier .

Both \ud835\udc5b and \ud835\udc65 are configurable parameters that we set at \ud835\udc5b = 50 and \ud835\udc65 = 60, parameters which were found via sensitivity analysis.

Results & Findings

A new generation of clinical decision support systems (CDSS) promises to reduce chances of delivering improper MV by automating aspects of ventilator configuration, and by providing clinically accurate and relevant alerts to providers. However, its use of a closed dataset, its limited temporal resolution, and the accuracy of the model represent potential limitations both for research and decision support , .

• A new generation of clinical decision support systems (CDSS) promises to reduce chances of delivering improper MV by automating aspects of ventilator configuration, and by providing.
• However, its use of a closed dataset, its limited temporal resolution, and the accuracy of the model represent potential limitations both for research and decision support.
• Having highly granular temporal resolution VM classification results is important because in practice providers may change VM frequently based on changes in clinical state or patient.
• To improve upon previous work, we note that machine learning (ML) has proven capable of accounting for the highly variable nature of physiologic data such as.
• So we created a ML model that could identify different VMs on a per-breath basis, with a freely accessible dataset, using only VWD as input.

Practical Applications

CDSS knowledge of VM is important because changing VMs may be a necessary procedure in the course of care for a patient . For example, if CDSS determines that a patient is breathing asynchronously with the ventilator, it may be able to make recommendation that providers choose a different VM that provides more comfort and flexibility in breathing to patients – .

Another example would be that CDSS could provide alerts to clinicians if patients continually violate safe volumes of air to inhale.

In this case the CDSS could recommend that patients be placed on a VM that limits tidal volumes such as volume-control.

Figures Explained

The paper’s visual material highlights the workflow and the main system components.

• Figure 1: Illustration of ventilation modes and their clinical relevance.. Highlights the importance of understanding ventilation modes for effective patient management.
• Figure 2: Performance of the model under varying levels of missing data.. Demonstrates the robustness of the model to incomplete data scenarios.

Conclusion

However, a key detriment to these systems is that they often lack access to the configured state of the ventilator and therefore lack information that may improve the efficiency of these CDSS . Therefore, one breath’s mode is often predictive of the next breath’s mode.

Using this methodology, we ablated the overall number of training observations by 71.41% from 140,928 to 40,285 observations, while still maintaining generalizability of our training set to our withheld test set, and largely improved CPAP performance (Table 4 ).

In conclusion, we created a highly-performant ML classifier for detecting five of the most commonly used ventilator modes in the USA, using only raw VWD as input.