Issue

Automated detection of asynchronisms during mechanical ventilation is an important and urgent task for the fastest recovery of patients with the least harm to the lungs.

For different  device’s modes, it is extremely difficult to calculate a universal algorithm that will allow you to detect asynchronisms. This is the main difficulty. At the same time, a person is able to visually identify various types of asynchronisms. Such problems are usually solved by machine learning methods, which we have implemented.

The key element of the Acrux Deep Breath ventilator monitoring system developed by Acrux Cyber ​​Services is the ML module, which allows to detect and classify asynchronisms in real time with high accuracy.

The collection and analysis of asynchronisms allows predicting the condition of patients, as well as selecting the optimal mode of operation.

Initial data evaluation

We had data for each patient's breath, as well as parameters for the inhalation process (50 measurements per second). As part of the initial data processing, our specialists visualized the data streams for breaths: indicators of air flow and pressure over time.

Visualization always helps to better understand the type of problem and choose the method by which to solve it. As you can see from the graphs above, it is impossible to find a simple formula that can be used to classify breaths.

Another problem was the number of such breaths - tens of thousands, and the difficulty of quickly classifying them manually to train the algorithms.

Solution

In scientific papers on this topic, we have seen the use of supervised learning methods. That is, labeled data (normal breaths, asynchronous) were loaded into the model and based on this data, the model learned to determine the classification of breaths. However, this task looked quite laborious, since one asynchronous inhalation accounts for 10-100 normal ones. Inspiratory classification is a task that cannot be delegated to unskilled people. Therefore, we took a different path.

The inspiratory data streams were divided into groups by the unsupervised learning method.

One of the intermediate results. One point - one breath. The closer the points are to each other, the more similar the breaths are to each other. Thus, groups of similar inspiration points are formed.

We managed to get 8 segments of breaths, which were then classified by specialists according to the types of asynchronism.

For example, this is the largest segment (more than 90% of breaths) are normal breaths.

Separating correct breaths from non-standard ones was the most important task that we managed to solve.

Example of inhalation from a cluster classified as ineffective triggering asynchronism

Other clusters and asynchronisms

Thanks to this separation, specialists have the opportunity not to waste time separating the correct breaths from the wrong ones. All attention was paid to non-standard cases of greatest interest. This analysis allowed us to correct machine learning algorithms and obtain a more accurate model.

Application

The identified asynchronisms are displayed in an easy-to-analyze view filtered by type and date.

This allows the clinician to adjust the ventilator settings to make ventilation safer for the patient.

When a sufficient number of successful solutions to the problem of asynchronisms have been collected, Acrux Deep Breath will be able to advise specialists on how to change the parameters in the optimal way. To do this, you will need to apply other ML methods.

Conclusion

In this article, we looked at how we managed to solve the problem of identifying and classifying asynchronisms using machine learning algorithms. Solutions didn’t require a lot of human’s time and allowed specialists to concentrate at the most difficult cases.

Similar methods are also used to detect fraudulent transactions, product defects and other unwanted events. We will consider other cases in the following articles. 

Subscribe to our social networks on LinkedIn and Facebook!