Deep Learning Could Aid Medical Professionals in Improved Medical Ventilator Control, Google AI Proposes
AI and Data Science NewsHealthcareGoogleposted by April Miller March 24, 2022 April Miller
Mechanical ventilators have gained a lot of publicity amid the COVID-19 pandemic. These medical devices are one of the most crucial tools hospitals have for supporting patients with breathing troubles, but running them effectively can be challenging. Google AI seems to have devised a solution.
In a recent study, Google AI researchers demonstrated how neural networks could control ventilators better than hand-tuned controllers. The machine learning algorithm analyzes air pressure to fine-tune the ventilator’s airflow to match specific patients’ breathing patterns. Early tests suggest it could outperform current manual methods and give medical staff more time.
Challenges With Current Ventilator Control Methods
To know how valuable machine learning controls can be, you first need to understand how ventilators work. Whereas normal breathing uses negative pressure to draw air in, these machines create positive air pressure to force air into patients’ lungs. It’s a potentially life-saving operation for those who can’t breathe normally, but it carries some significant risks.
Ventilators must deliver a precise waveform to avoid putting too much or too little air into patients’ lungs. Since every patient has unique respiratory systems, that balance changes from person to person. As a result, operating a ventilator requires continuous attention from experienced doctors to adjust the waveform and prevent lung damage.
Today’s ventilators use PID (Proportional, Integral, Differential) controls, which adjust the device based on discrepancies between the measured and target air pressures. It’s the same kind of system you’ll find in an espresso machine. These readings provide a reliable baseline, but they’re prone to over and under-shooting targets because of differences between patients, requiring continuous monitoring.
Google’s Ventilator Control Solution
Google’s alternative uses machine learning instead of manual controls. The researchers started with a physical test lung and an open-source ventilator design to run simulations. A series of traditional PID controller simulations on various lug settings provided a baseline, and then Google AI trained a deep neural network (DNN) to monitor and predict airway pressure.
Since the DNN uses patient-specific, real-time data to predict air pressure changes, it can prevent the dramatic corrections you see in PID controllers. It adjusts to prevent drastic pressure changes instead of responding to them.
The DNN showed a 20% lower mean absolute error (MAE) between target and actual air pressure than the best-performing PID controller. It also had a 32% lower MAE when working across multiple lung conditions, suggesting this method requires far less manual intervention if a patient’s condition changes.
The reduced need for doctor monitoring and intervention is crucial. Healthcare staff shortages will likely continue for five to ten years, with 6.5 million employees leaving their positions by 2026 and only 1.9 million coming in to replace them. Hospitals need to address more patients with less staff, and freeing doctors’ time with automated ventilator controls will do that.
Automating the adjustment process also helps reduce the risk of human error. Medical malpractice is the third leading cause of death in the U.S., and even a small misunderstanding can cause serious damage. This DNN’s accuracy is a promising sign that automation could lead to more accurate treatments.
While this DNN-controlled ventilator concept is certainly promising, some roadblocks remain. As the Google AI researchers point out, the tests cover various lung conditions but aren’t entirely representative. The researchers didn’t train or test the DNN on prenatal, child, or non-sedated patients, so it may not be effective in those situations.
The tests also only looked at invasive ventilation. Since the masks in non-invasive ventilation create their own positive pressure, they could skew the results. DNNs may have difficulty telling the difference between pressure from the mask and pressure from the airway. Those kinds of nuances are still challenging for machine learning algorithms, so DNN-controlled ventilators may only work in invasive operations.
As with many machine learning applications, this DNN may not be able to account for all real-world variables it could encounter. Outside of the simulations, there could be far more unexpected factors to account for, which the neural network may struggle with.
Machine Learning Could Save Lives
While significant challenges remain, Google AI’s research paints a bright picture of the future. With some more tests, training, and real-world applications, machine learning algorithms could help automate ventilators and other medical equipment.
Automating medical devices through DNNs could help reduce hospital staff stresses and the complications they bring. Overall patient care could improve as a result, saving lives in the end.