This study analyzed the characteristics of 4747 breaths recorded on patients who were ventilated with a single limb circuit with a leak port. First, we categorized the breaths into three types based on the CO2 motion. Second, we demonstrated that simple characteristics that can be measured on pressure and flow curves differ according to the three types of CO2 motion. Third, we showed that we could use a deep learning technique, multilayer perceptron (MLP) network, to recognize each type of breath. The results suggest that by analyzing and calculating parameters that are routinely measured by modern ventilators, it is possible to monitor the type of every breath, thereby allowing to adjust the ventilator settings to minimize the risk of CO2 rebreathing. With the computing capabilities of modern ventilators it should be easily calculated in real time to warn the medical staff of the increased risk of CO2 rebreathing.
Importantly, this study took into account the complexity of the actual continuous interactions between the effects of mechanical ventilation and their effects on the physiological regulation of ventilation in terms of drive, frequency, and the inspiratory/expiratory ratio and the possibility of non-intentional leaks. We should emphasize how the calculations were performed. At first, each breath type was identified “visually” according the CO2 motions described at Fig. 2. Then we found that the main characteristics showed significant differences according to the type of breath, allowing us to correctly differentiate each type (with 97.9% accuracy for type III breaths). Finally, we used a deep learning method to analyze and categorize the three types of breath.
Our sample comprised patients who were intubated or who were non-invasively ventilated via a mask that covered the mouth and nose. The distribution of breath types in the intubation and noninvasive ventilation groups, respectively, were as follows: type I, 64% and 36%; type II, 66% and 34%; and type III, 74% and 26%. It was possible that the process needed to identify the breath type might differ according to intubation versus mask ventilation status; however, since the identification was 100% accurate using the MLP network, this does not appear to be a concern. Indeed, this method seems to be usable regardless of whether the patient is intubated or is using an oronasal mask.
Another possible concern is that the type of breath could always be the same for a particular patient using specific ventilator settings, making continuous analysis not very relevant. However, we found that for each patient, the average percentage of breaths that showed significant rebreathing (FICO2 > 0.1%) was 45% ± 31%. This clearly indicates that each patient shows both rebreathing and non-rebreathing. This is illustrated on the patients’ records, some of which show the three types of breath during the same minute. In a previous study [8], logistic regression analysis showed that rebreathing may depend both on the previous value of the end tidal level of CO2 (OR = 3.09) and on the instantaneous respiratory rate expressed as breaths per minute (OR = 1,19). In that study, a leak valve was mandatory with the single limb circuit, which was the classical Respironics Whisper Swivel that is located on the circuit close to the intubation tube or to the mask. Some studies have shown that if there is a noninvasive interface, a leak incorporated into the mask itself decreases rebreathing in bench studies [6, 7]. Of course, this makes breathing more efficient by decreasing the dead space of the mask and thus the VD/VT ratio. Nevertheless, the mechanism and the risk of rebreathing we describe here remain the same.
Some studies propose that a plateau exhalation device or a non-rebreather valve be used to limit the risk of rebreathing and show that these kinds of devices can eliminate rebreathing at the expense of an increase in work of breathing and of expiratory resistance [9]. On the other hand, in patients with COPD, rebreathing increases during exercise; this is related to the high expiratory flow, which is not totally eliminated during shortened expiratory duration, which may increase in the respiratory frequency [10]. Notably, non-rebreather valves are not currently recommended, especially since a clinical trial showed negative results [9].
One limitation of our study was that the method we described here allows only a qualitative measure of whether there is rebreathing; it does not actually quantify rebreathing. Doing so would mainly be important if a lot of the breaths were type III. It is also possible to identify type II breaths, which have a higher risk of rebreathing.
Another key point to keep in mind is the potential clinical importance of the amount of rebreathing that comes from the circuit and adds to the anatomical dead space ventilation that we identified. Some studies note that this is not a cause for concern, especially in noninvasive ventilation, due to non-intentional leaks [11]. Nevertheless, the impact of even a slight increase or decrease in PaCO2 and the impact of frequent and rapid changes in PaCO2 remain unknown. However, a change as small as 1 mmHg in the PaCO2 is a powerful stimulus in terms of regulating respiratory drive, and the PaCO2 level has important hemodynamic effects that are mediated mainly by sympathetic stimulation of pulmonary tension, peripheral tension, and cerebral blood flow [12–14 ]. On the other hand, clinical studies clearly show that in either acute or chronic hypercapnic respiratory failure, the aim of ventilator assistance is to decrease the PaCO2 level, which may be impeded if the BiPAP circuit induces rebreathing.