Decreased breathing variability is associated with poorer outcome in mechanically ventilated patients

Rationale Breathing is a cyclic activity that is variable by nature. Breathing variability is modified in mechanically ventilated patients. We aimed to evaluate whether decreased variability on the day of transition from assist-control ventilation to a partial mode of assistance was associated with a poorer outcome. Methods This was an ancillary study of a multicentre, randomised, controlled trial comparing neurally adjusted ventilatory assist to pressure support ventilation. Flow and the electrical activity of the diaphragm (EAdi) were recorded within 48 h of switching from controlled ventilation to a partial mode of ventilatory assistance. Variability of flow and EAdi-related variables were quantified by the coefficient of variation, the amplitude ratio of the spectrum's first harmonic to its zero-frequency component (H1/DC) and two surrogates of complexity. Main results 98 patients ventilated for a median duration of 5 days were included. H1/DC of inspiratory flow and EAdi were lower in survivors than in nonsurvivors, suggesting a higher breathing variability in this population (for flow, 37% versus 45%, p=0.041; for EAdi, 42% versus 52%, p=0.002). By multivariate analysis, H1/DC of inspiratory EAdi was independently associated with day-28 mortality (OR 1.10, p=0.002). H1/DC of inspiratory EAdi was lower in patients with a duration of mechanical ventilation <8 days (41% versus 45%, p=0.022). Noise limit and the largest Lyapunov exponent suggested a lower complexity in patients with a duration of mechanical ventilation <8 days. Conclusion Higher breathing variability and lower complexity are associated with higher survival and lower duration of mechanical ventilation.


Introduction
Breathing is a cyclic activity that is not monotonous, but exhibits natural variability [1,2]. In normal human subjects, ventilation shows breath-by-breath variability in descriptors of breathing pattern such as respiratory rate and tidal volume [2]. Breathing variability can also be characterised by the spectral analysis of the flow signal [3,4]. Finally, breathing activity is nonlinear in nature and exhibits chaos-like mathematical complexity [5,6]. as a prognostic index at a given time point of ICU stay. In addition, in these studies, breathing variability was restricted to downstream variables such as airway flow and tidal volume, while the upstream variability of the central inspiratory activity was ignored. It is worth noting that the electromyographic activity of the diaphragm depends directly on the central inspiratory activity [8]. Finally, variability was generally assessed using one single tool of analysis. A recent review describes breathing variability in anaesthesia and critical care, suggesting that variability of respiration is not yet fully understood and that the respiratory system should be measured as a whole rather than a single parameter [9].
Here, we performed an ancillary study of a multicentre, randomised, controlled trial. We described and quantified variability by an array of descriptors, including breath-by-breath variability, spectral analysis and mathematical complexity. This quantification was achieved at the transition between assist-control ventilation and ventilation with a partial mode; in other words, as soon as patients could sustain pressure support ventilation (PSV). We chose this time point because this is the first moment during mechanical ventilation that the brain resumes its control over ventilatory activity and therefore the first moment that the natural variability of the respiratory system can be evaluated, since this natural variability was previously occulted by control ventilation [10]. We hypothesised that a low variability at the time of the switch to partial ventilatory mode could predict a poorer outcome.

Methods
This is a pre-planned ancillary study of a multicentre, randomised, controlled trial that aimed to compare neurally adjusted ventilatory assist (NAVA) to PSV in mechanically ventilated patients in 11 ICU departments in France [11]. The study protocol was approved for all centres by the Comité de Protection des Personnes Ile de France 8 (no. 2010-A00424-35), according to French law. A detailed description of the study design and data from this cohort has been published previously [11,12].

Patients
Patients receiving mechanical ventilation for >24 h for acute respiratory failure of respiratory cause were eligible when they met the following criteria: ability to sustain PSV for ⩾30 min with a total level of inspiratory pressure <30 cmH 2 O, estimated remaining duration of mechanical ventilation >48 h, level of sedation (Ramsay scale) ⩽4, fraction of inspired oxygen ⩽50% with a positive end-expiratory pressure ⩽8 cmH 2 O and absence of administration of high-dose vasopressor therapy. Exclusion criteria were age <18 years, known pregnancy, participation in another trial within the 30 days preceding satisfaction of the eligibility criteria, contraindication of the implementation of the oesophageal tube and decision to withhold life-sustaining treatment.

Patient management
After inclusion, patients were connected to a Servo-i ventilator (Maquet Critical Care, Sweden) equipped with NAVA mode. An extensive description of patient management is provided in the supplementary methods.

Data collection
Airway pressure, airway flow and the electrical activity of the diaphragm (EAdi) were recorded 12, 24, 36 and 48 h after inclusion. They were acquired for 20 min at 100 Hz from the ventilator connected to a computer using commercially available software (Servo-i RCR, version 3.6.2; Maquet Critical Care). Outcome data included mortality 28 days after inclusion, duration of mechanical ventilation and ventilator-free days (VFDs) 28 days after inclusion.

Data analysis
For each patient, the four 20-min recordings (12,24,36 and 48 h after inclusion) were merged into one single 80-min recording session on which analyses were performed ( figure 1). An extensive description of signal processing and data analysis is provided in the supplementary methods.
Breath-by-breath variability of flow-derived and EAdi-derived breathing pattern variables was assessed by the coefficient of variation (standard deviation divided by the mean; the higher the coefficient of variation, the higher the variability). Flow-derived breathing pattern variables included tidal volume and respiratory rate. For EAdi, peak EAdi (EAdi-peak) and EAdi-inspiratory neural time were determined.
Spectral-derived variability was assessed using the amplitude ratio of the spectrum's first harmonic (H1) to its zero-frequency or DC component (H1/DC) according to the method described by GUTIERREZ et al. [4] (the higher the H1/DC, the lower the variability).
Breathing complexity was assessed by the noise limit and largest Lyapunov exponent [13]. A noise limit above zero means nonlinearity and a certain degree of complexity [10,14,15]. Sensitivity to initial conditions is how perturbations occurring in the past affect the future behaviour of the system and is another characteristic of how a complex system is unpredictable. This was estimated for flow and EAdi using the largest Lyapunov exponent [13].

Statistics
As this is an ancillary study, no sample size could be calculated to detect a difference. The sample size was determined by the parent study [11]. Statistical analysis was performed using GraphPad (GraphPad Software, San Diego, CA, USA) and R (The R Foundation, Vienna, Austria). Continuous data were reported as median (interquartile range) and categorical data as number of events ( percentage). Continuous variables (i.e. duration of mechanical ventilation and number of 28-day VFDs) were dichotomised according to their median value in the population. The four recordings merged into one single 80-min recording Transition from assist-control ventilation to a partial mode of assistance Supplementary table SDC1 compares the descriptors of breathing variability between patients assigned to the NAVA group and those assigned to the PSV group in the mother trial. The coefficient of variation of the tidal volume and the largest Lyapunov exponent for flow were higher in patients assigned to the NAVA group as compared to those assigned to the PSV group.
Association between breathing variability and 28-day mortality Mortality within 28 days was 19% (n=19). Table 2 shows the descriptors of breathing variability associated with 28-day mortality by univariate analysis. Among descriptors of breathing variability, two differed between survivors and nonsurvivors. H1/DC of inspiratory flow and H1/DC of inspiratory EAdi were lower in survivors, suggesting a higher variability in this population. By multivariate analysis, H1/DC of inspiratory EAdi was the only factor independently associated with 28-day mortality (OR 1.10, 95% CI 1.04-1.17; p=0.002).
Association between breathing variability and duration of mechanical ventilation Duration of mechanical ventilation was 8 (4-13) days. Table 3 shows the descriptors of breathing variability associated with duration of mechanical ventilation. Among descriptors of breathing variability, three differed between patients with a duration of mechanical ventilation <8 days and those with a duration of mechanical ventilation ⩾8 days. H1/DC of inspiratory EAdi was lower in patients with a duration of mechanical ventilation <8 days, and there was a significant, but poor, correlation between H1/DC of inspiratory flow and EAdi and duration of mechanical ventilation (supplementary figure SDC1). This suggested a higher breathing variability in patients with a shorter duration of mechanical ventilation. Noise limit for respiratory flow and EAdi was higher in patients with a longer duration of mechanical ventilation, and there was a positive correlation between noise limit for respiratory flow and EAdi and duration of mechanical ventilation. This suggested an association between a higher complexity and a longer duration of mechanical ventilation (supplementary figure SDC1, supplementary table SDC2).

Association between breathing variability and 28-day VFDs
Ventilator-free duration 28 days after inclusion was 23 (10-25) days. Table 4 shows the association between descriptors of breathing variability and 28-day VFDs. Among descriptors of breathing variability, eight differed between patients with 28-day VFDs <23 days and those with 28-day VFDs ⩾23 days.
Among patients with 28-day VFDs ⩾23 days, the coefficient of variation of the tidal volume was higher and the inspiratory and expiratory H1/DC for EAdi and flow were lower, suggesting an association between a higher variability and an increased number of 28-day VFDs. Correlations between these variables and 28-day VFDs conveyed the same message. Among patients with 28-day VFDs <23 days, the noise limit of flow and EAdi and largest Lyapunov exponent of flow were higher, with a correlation between these variables and 28-day VFDs. These results suggested that a higher complexity was associated with fewer 28-day VFDs ( figure 2 and supplementary table SDC3).

Discussion
The main findings in our cohort of 98 mechanically ventilated patients studied at the early phase of weaning can be summarised as follows: 1) higher breath-by-breath variability as assessed by the coefficient of variation, and higher spectral variability as assessed by the H1/DC ratio, were associated with a lower mortality and a lower duration of mechanical ventilation, resulting in increases in VFDs; 2) higher complexity as assessed by noise limit and the largest Lyapunov exponent was associated with a longer duration of mechanical ventilation and fewer VFDs.
To our knowledge, this is the first study to evaluate in a large population the prognostic impact of reduced variability and complexity in the ICU at one given time point (i.e. the transition between assist-control ventilation and a partial mode of assistance such as PSV or NAVA) and with several flow-and EAdi-derived indices to quantify breath-by-breath variability, spectral-derived variability and complexity. Previous studies on this topic used only one of these approaches and did not integrate EAdi into their analyses.

Relationship between variability and outcome
A major result was that higher breath-by-breath and spectral variability were associated with a better outcome. This result is in line with previous reports showing that a higher breath-by-breath variability is associated with a higher weaning success rate [7,16]. A body of literature suggests an inverse relationship between breathing variability and respiratory system loading [17][18][19]. In mechanical ventilation patients, unloading the respiratory system is associated with higher respiratory variability [20,21]. These results suggest that respiratory variability parallels the load-capacity balance of the respiratory system. A high variability may witness a large respiratory reserve [7], and subsequently a higher likeliness to be weaned with, in turn, a shorter duration of mechanical ventilation [22]. Regarding the association between higher spectral variability as assessed by the H1/DC ratio and lower mortality, our findings confirm the previous report from GUTIERREZ et al. [3].

Relationship between complexity and outcome
Greater complexity was associated with a longer duration of mechanical ventilation and fewer VFDs. Ventilatory flow is not periodic [5], but exhibits complexity, with this term implying irregularity, sensitivity to initial conditions and unpredictability. In other words, this is the amount of "surprise" or "new information" introduced into an otherwise predictable system, i.e. the degree of disorder or randomness in the data [9]. In animals and humans, ventilator complexity has been characterised by various mathematical approaches such as correlation dimension, approximate entropy, Lyapunov exponents and noise limit, which investigates the chaotic nature of ventilator flow [23].
Few studies have evaluated the relationship between complexity and outcome in ICU patients. These studies are in line with our results. The study by EL-KHATIB et al. [24] showed that the breathing pattern measured by Kolmogorov entropy and respiratory flow-volume phase space dimension during mechanical ventilation was more complex and chaotic in patients who failed weaning than in those who succeeded. The study by ENGOREN et al. [25] found similar results. Patients who failed weaning showed increased irregularity in the biosignal analysis of approximate tidal volume entropy, which, according to the authors, reflected enhanced external inputs to the respiratory control centre. They suggested that increased regularity in the weaning success group indicated a better adaptive mechanism of an autonomous system. Finally, PARK et al. [26] found that the electrocardiogram and photoplethysmography exhibited more complex and chaotic behaviour in patients who failed weaning.

Clinical implications and perspectives
Our results suggest that breathing variability measured at a given time point, the transition between assist-control ventilation and a partial mode of assistance, could be used as a predictor of duration of mechanical ventilation and even survival. This may help in deciding important therapeutic options such as hastening the weaning process or, conversely, performing a tracheostomy. It is worth noting that analyses derived from the upstream EAdi signal did not provide much more information than analyses derived from the downstream flow signal, which will simplify the assessment of variability in daily practice, since recording the respiratory flow signal is much easier than recording the EAdi. This result was quite surprising, since EAdi is a closer surrogate of the activity and hence variability of the central respiratory pattern generators located in the brainstem [8]. It suggests that the prognostic value of breathing variability results not only from the central respiratory pattern generator from where it originates [23], but also from the way the respiratory system alters this neural variability, which relates to the load-capacity relationship of the respiratory system [17,18,27].
Because greater variability is associated with a better outcome, one can hypothesise that restoring variability could improve the outcome. A body of literature suggests that, during mechanical ventilation, greater variability may be associated with a more protective ventilation. In mechanically ventilated animals, decreased variability of tidal volume is associated with altered lung mechanics and increased lung damage [28], and the restoration of a certain level of variability [29][30][31] improves respiratory system compliance and the secretion of surfactant, decreases histological lung damage and lung inflammation and improves gas exchange [28][29][30][31][32]. Restoring variability could involve the restoration of the intrinsic variability of the respiratory system with a proportional mode of ventilation such as NAVA or proportional assist ventilation [8,33]. Previous studies have shown that breath-by-breath variability is higher with these modes than with pressure support ventilation [8,33]. This could involve the introduction of a certain level of extrinsic variability with modes of mechanical ventilation such as variable or "noisy" pressure support ventilation [34,35]. In mechanically ventilated patients, this mode is associated with improved gas exchange [22].
From a clinical perspective, our results pave the way for future studies evaluating how breathing variability could be used to improve the management of mechanical ventilation. For instance, combined with other anamnestic or clinical data, breathing variability may help to determine the outcome of a patients transitioning from assist-control ventilation to pressure support.
In the era of artificial intelligence and personalised medicine, our results could be later used as a predictive algorithm for weaning success or failure and to adjust the promptness of transition from controlled and assist-control ventilation to a partial mode of assistance. In addition, mechanically ventilated patients at high risk of mortality will be more easily identified.

Strength and limitations of the study
The strengths of this study include the unselected character of our population of ICU patients, which is quite representative of a standard ICU population given its characteristics, severity and outcome. The multicentre design, involving 11 ICUs, enhances the generalisability of our findings. Finally, all the patients were studied at a given and comparable time point. This study presents a number of limitations. First, the sample size was not calculated a priori because it was a secondary analysis. Second, the recordings could not be analysed in some patients, which reduced the sample size and in turn decreased the power of the study. Third, some of the indices we used required long and complex mathematical processing, which limits the immediate transposition of our results. Fourthly, aggregating measurements made over 48 h could "dilute" the moment when the brain recovers its aptitude to generate variability. However, limiting the analysis to the first recording would have limited the quality of analyses due to the short duration (20 min) of the recording. Finally, our study suggests how to monitor the transition from assist-control ventilation to a partial mode of assistance in a large but heterogeneous population and confounders as disease severity, comorbidities, baseline diagnostics may have impacted the results. Further studies are therefore needed to determine in more balanced groups the impact of our measurements. In this preliminary, and by no means exhaustive study on the use of an array of variability and complexity descriptors, it would be nice to further compared the analysed parameters between the different weaning groups (i.e. short, difficult and prolonged weaning) keeping only the analysis of the significant parameters identified in this work.

Conclusion
In mechanically ventilated patients studied at the transition from assist-control ventilation to a partial mode of assistance, higher breath-by-breath variability and spectral variability were associated with better outcomes. These results pave the way for future studies that will evaluate more precisely the accuracy of these indices, which time point is the more reliable to gather them, and if repeated measures could improve this accuracy. Obviously, these studies will require the development of automated tools. In addition, these results support trials that would evaluate the prognostic impact of strategies aiming at restoring a more physiological level of variability in mechanically ventilated patients, although this physiological level is as yet unknown [2,36].