We implemented and tested four different techniques in the system described above. Our two best techniques were based on modifying the AIC techniques to create an adaptive physiologic model of the lungs. The second technique was based on using 3 different adaptive physiologic models in a mixture of experts system (switching based on predictive performance of the models).

In pressure support ventilation, the most appropriate performance measure is work of breathing (WOB). The goal of pressure support ventilation is to consistently reduce the WOB of the patient, without using excessively high pressures to achieve this goal. When determining the WOB of the patient, the standard method is to compute the overall WOB and subtract off the ventilator WOB. It is relatively simple to compute the WOB done by the ventilator, but much more difficult to compute the overall WOB. For our study, however, it is sufficient to say that increased ventilator WOB means decreased patient WOB. Thus, the models with the most ventilator WOB are the best models.

The ventilator we compared our AIC control systems to was the Puritan-Bennett 7200 ventilator. This is the most common US intensive care unit (ICU) ventilator and costs between $25,000 and $50,000. A few important issues must be considered. First, commercial ventilators have multiple pressure sensors and a flow sensor, whereas we only have one pressure sensor. Second, commercial ventilators are very carefully calibrated and maintained, ours is not. Third, the control algorithms in commercial ventilators are very highly tuned to the particular task they are implementing. With all of this in mind, two of our methods outperformed the PB7200 when compared using work of breathing.

Table 1 shows the qualitative results over different lung compliances and pressure targets. Notice that in all cases both our physiologic model and predictive switching mixture of experts system outperformed the PB7200. For the physiologic model, the average increase in ventilator work of breathing was 24% more than the PB7200 and was never less than 17% better. The predictive switching mixture of experts system improved on the physiologic model by approximately 4% on average.

Again, remember that we are comparing our prototype system with one sensor against a highly tuned, calibrated $50,000 ventilator with multiple sensors. Clearly these results are significant.

CLICK ON THE INDIVIDUAL VALUES TO SEE A PLOT COMPARING THE PRESSURE WAVEFORMS

* PB = Puritan-Bennett 7200 ventilator, RC = adaptive physiologic model, ME = adaptive mixture of experts

Ventilator Home | Next Page

Products | Support | Order | Download | Search | Contact

Product questions? Contact info@nd.com Website questions? Contact webmaster@nd.com
Web Site Design and Implementation Copyright © 2002 NeuroDimension, Inc.