Authors: Alison Wallbank (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.), Alexander Sosa (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.), Andrew Colson (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.), Huda Farooqi (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.), Elizabeth Kaye (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.), Katharine Warner (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.), David J. Albers (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.; 2Department of Biomedical Informatics, University of Colorado Anschutz Medical Campus, Aurora, CO 80045, USA), Peter D. Sottile (3Division of Pulmonary Sciences and Critical Care Medicine, Department of Medicine, University of Colorado School of Medicine, Aurora, CO 80045, USA), Bradford J. Smith (1Department of Bioengineering, University of Colorado Denver | Anschutz Medical Campus, Aurora, Colorado, USA.; 4Section of Pulmonary and Sleep Medicine, Department of Pediatrics, University of Colorado Denver, Anschutz Medical Campus, Aurora, CO 80045, USA)
Categories: Article, acute lung injury, mechanical ventilation, ventilator-induced lung injury, driving pressure, mechanical power
Source: American journal of physiology. Lung cellular and molecular physiology
Authors: Alison Wallbank, Alexander Sosa, Andrew Colson, Huda Farooqi, Elizabeth Kaye, Katharine Warner, David J. Albers, Peter D. Sottile, Bradford J. Smith
Mechanical ventilation is a necessary lifesaving intervention for patients with Acute Respiratory Distress Syndrome (ARDS) but it can cause ventilator induced lung injury (VILI), which contributes to the high ARDS mortality rate (≈40%). Bedside determination of optimally lung-protective ventilation settings is challenging because the evolution of VILI is not immediately reflected in clinically available, patient-level, data. The goal of this work was therefore to test ventilation waveform-derived parameters that represent the degree of ongoing VILI and can serve as targets for ventilator adjustments. VILI was generated at three different positive end expiratory pressures in a murine inflammation-mediated (lipopolysaccharide, LPS) acute lung injury model and in initially healthy controls. LPS injury increased expression of proinflammatory cytokines and caused widespread atelectasis, predisposing the lungs to VILI as measured in structure, mechanical function, and inflammation. Changes in lung function were used as response variables in an elastic net regression model that predicted VILI severity from tidal volume, dynamic driving pressure (PDDyn), mechanical power calculated by integration during inspiration or the entire respiratory cycle, and power calculated according to Gattinoni’ s equation. Of these, PDDyn best predicted functional outcomes of injury using either data from the entire dataset or from 5-minute time windows. The windowed data shows higher predictive accuracy after a ≈1-hour ‘run in’ period and worse accuracy immediately following recruitment maneuvers. This analysis shows that low driving pressure is a computational biomarker associated with better experimental VILI outcomes and supports the use of driving pressure to guide ventilator adjustments to prevent VILI.
Mechanical ventilation (MV) is a necessary, life-saving intervention to treat patients with acute respiratory distress syndrome (ARDS). The clinical standard for MV employs the ARDSnet protocol of small tidal volume (Vt) breaths supported by incrementally increasing positive-end expiratory pressure (PEEP) (1–4). Despite the improvement in survival associated with this approach, ventilator-induced lung injury (VILI) often still occurs and contributes to the high (≈40%) ARDS-associated mortality rate (1, 3, 5).
One of the key challenges to reducing VILI and improving ARDS outcomes is that lung injury has a unique effect on the lungs of each patient due to differences in injury etiology and the patients themselves. This is compounded by temporal changes in injury severity and pathophysiology so that there is no ‘universal’ lung protective setting that is always best for all patients. Instead, the ventilator must be repeatedly adjusted to match individual pathophysiology and maintain adequate gas exchange while avoiding VILI, which raises the challenge of determining how to guide ventilation to optimize lung protection.
Conceptually, the goal of this study is to test what we term VILI cost functions (6, 7) that describe the severity of ongoing VILI during MV so that ventilator settings may be adjusted to minimize that value and protect the lung. Given the real-time availability of respiratory data from the ventilator, and the association between respiratory mechanics and VILI susceptibility (1, 3), cost functions derived from pressure-volume data or ventilator parameters are an alluring solution. Tidal volumes are an example where a respiratory parameter is monitored and ventilation is adjusted to meet a goal, and provide a striking example of how this approach can improve outcomes (8). However, tidal volumes do not account for inter-patient differences in the response to ventilation where, for example, higher tidal volumes are more injurious in pre-injured lungs (9, 10). Several alternative cost functions have been proposed and tested, including the dynamic driving pressure and different forms of mechanical power. These seek to incorporate the effects of both ventilator settings, including Vt and the positive end expiratory pressure (PEEP), the underlying mechanical function of the lungs, and the interaction between the lungs and the ventilator; for example, the reduction in elastance that occurs due to recruitment after selection of an appropriate PEEP. However, questions remain about the efficacy and applicability of these approaches.
To experimentally test the association between these VILI cost functions and experimental outcomes, mice with and without endotoxin-induced acute lung injury (ALI) were ventilated with PEEP = 1, 3, or 8 cmH2O for four hours. The effects of ALI and VILI were measured in lung mechanical function, pro-inflammatory gene expression, alveolocapillary barrier injury, and lung structure (stereology). The VILI cost functions, including the dynamic driving pressure and different forms of mechanical power, were computed from continuous measurements of pressure and flow at the airway opening. An elastic net regression was used to determine the most important predictors of VILI, as measured in lung function, using data from the entire ventilation period or from 5-minute time windows over the course of the experiment.
All animal studies described were approved by the Institutional Animal Care and Use Committee (IACUC, protocol #00230) at the University of Colorado Denver | Anschutz Medical Campus. All mice were acclimated for at least one week prior to experimentation.
Eight- to ten-week-old C57BL/6J female mice (Jackson Laboratory) were injured intratracheally (IT) with lipopolysaccharide (LPS, L4524, MiliporeSigma). Each mouse received 50μg LPS dissolved in 50μL sterile phosphate buffered saline (PBS) at time zero. Control (CTL) mice did not receive an instillation. We elected not to employ a vehicle (PBS) instillation in the control group because we are not attempting to deconvolve the effects of endotoxin from the potentially injurious PBS instillation. Instead, we sought to analyze VILI in both lung-injured (LPS) and initially healthy (CTL) mice as these two groups were hypothesized to have the broadest divergence in their response to ventilation. Forty-eight hours after IT injury with LPS, mice were removed from housing and transported to the laboratory for lung function assessment, mechanical ventilation, and tissue harvest as shown in Figure 1.
Mice from each injury group were randomized into four subgroups. Non-ventilated mice were designated NV. NV mice harvested for biochemical analysis were sacrificed and harvested directly following deep anesthesia. NV mice designated for morphometric analysis were briefly placed on the mechanical ventilator for the initial phases of the tissue preparation described below.
The ventilated cohort of animals were deeply anesthetized with acepromazine (2.5mg/kg), ketamine (100mg/kg), and xylazine (8mg/kg) via intraperitoneal (IP) injection, tracheostomized with a blunted 18-gauge (ga) thin-walled metal canula, and mechanically ventilated (SCIREQ flexiVent). Respiratory drive was suppressed with 0.8mg/kg IP pancuronium bromide. Each mouse underwent 10 minutes of stabilizing ventilation with tidal volume (Vt) = 10mL/kg, respiratory rate (RR) = 150 breaths per minute, and positive end expiratory pressure (PEEP) = 3 cmH2O with recruitment maneuvers (RM, a 3-second ramp to 30 cmH2O with a 3-second hold) every two minutes.
A comprehensive lung function assessment was then conducted. First, a RM was performed followed by a 16-second stepwise pressure-volume (PV) loop with an onset pressure of 0 cmH2O and a maximum pressure of 30 cmH2O to calculate inspiratory capacity (IC, the delivered volume) and quasi-static compliance (Cst, the slope of the quasi-static points at 5 cmH2O on the expiratory limb). The PEEP was then set to 9 cmH2O, a RM was applied, and four 3-second multi-frequency forced oscillations were recorded at 10 second intervals to determine pulmonary system elastance (H), tissue damping (G), and airway resistance (Rn) by fitting to the constant phase model (11). That measurement sequence was then repeated at PEEP = 6, 3, and 0 cmH2O.
The ventilated cohorts each received 8.5 ml/kg delivered tidal volume with PEEP = 1, 3, or 8 cmH2O (PEEP1, PEEP3, and PEEP8 groups, respectively) for four hours. The comprehensive mechanics scan was then applied again to assess ventilator-induced changes in lung function. Deep anesthesia was maintained throughout the ventilation period with alternating IP injections of ketamine and a ketamine/xylazine mixture every forty minutes. IP sodium bicarbonate (0.84% with a volume of 10μL/g bodyweight) was delivered with each drug administration. Separate cohorts of ventilated mice were prepared for morphometric and biomarker analysis.
A custom pneumotachometer with a resistance ≈ 0.2 cmH2O/ml/s was positioned downstream of the ventilator wye tubing. The differential pressure, which is proportional to flow rate, and the airway pressure were recorded at 256 Hz using an Arduino microcontroller coupled to a MATLAB graphical user interface (GUI). The GUI also proved closed-loop control of tidal volume using the pneumotach readings and programmatic interaction with the flexiVent control software (flexiWare).
Mice were randomized into non-ventilated and ventilated groups. NV mice were anesthetized (above), tracheostomized, and paralyzed with 0.8mg/kg IP pancuronium bromide. Bronchoalveolar lavage fluid (BALF) was collected by instilling and gently suctioning 0.8 mL ice-cold PBS twice through the tracheal cannula and centrifuging at 1000 G for five minutes. The supernatant was aspirated and stored at −80°C. Whole lungs were then excised, separated from lymphatic and cardiac tissue, snap-frozen with liquid nitrogen, and stored at −80°C.
Supernatant from the BALF was initially analyzed with the Thermo Scientific Pierce BCA Protein Assay Kit according to the manufacturer’s instructions. Processed samples were read at 562 nm with a spectrophotometer to assess total BALF protein content. Samples were run in triplicate. Once the total protein concentration of the BALF supernatant was analyzed with the BCA, four colorimetric sandwich ELISAs were performed in duplicate using the BALF supernatant for albumin, IL-6, TNF-α, and CXCL-1 according to manufacturer’s instructions. Albumin concentration was determined at a 600 dilution with the Bethyl Laboratories Mouse Albumin ELISA Kit, IL-6 concentration was measured at a 2 dilution with the Invitrogen Mouse IL-6 Uncoated ELISA kit, and both TNF-α and CXCL-1 were analyzed without dilution with the Invitrogen Mouse TNF alpha Uncoated ELISA kit and Invitrogen Mouse CXCL1 (KC) ELISA kits. The samples were read at 450 nm with a microplate reader to assess the respective protein content.
Whole lung tissue was analyzed with real-time quantitative polymerase chain reaction (RT-qPCR) for relative expression of IL-6, CXCL-1, TNF-α, and NF-κB (n=5 per group) per manufacturer instructions. Briefly, the whole lungs were homogenized in 2 mL Qiazol reagent using a gentleMACs tissue dissociator and then centrifuged at 400 G for 7 min. The supernatant was collected and transferred to tubes containing 0.4 mL chloroform and incubated for three minutes. Following incubation, the tubes were centrifuged at 12,000 G for 15 min at 4°C, yielding RNA contained in aqueous phase. Samples were stored at −80°C until RNA analysis.
Isolated RNA was converted to cDNA (Applied Biosystems RT kit) which was amplified by reverse transcriptase amplification with the BioRad CFX-9600 thermal cycler. RNA was quantified using nano-drop and normalized to 125 ng/uL with a final mass of 250 ng per reaction. RT-qPCR was performed for proinflammatory genes IL-6, CXCL-1, TNF-α, and NF-κB with glyceraldehyde 3-phosphate dehydrogenase (GAPDH) used as the housekeeper gene for normalization per manufacturer’s instructions. ThermoFisher Taq primers for IL-6, CXCL-1, TNF-α, NF-κB, and GAPDH were used.
A bilateral thoracotomy was performed on mechanically ventilated mice, the inferior vena cava bisected, and the pulmonary circulation was flushed with heparinized saline containing 3% 100 kDa dextran at an instillation pressure of 35 cmH2O through the right ventricle. The lungs were then inflated to an airway pressure of 30 cmH2O, ramped down and held at 5 cmH2O, and the trachea was ligated. The pulmonary vasculature was then perfused with fixative (4% paraformaldehyde, 1% glutaraldehyde, 3% 100 kDa dextran in 0.15 M HEPES buffer) at 35 cmH2O for ≈5 minutes. The heart-lung block was excised and immersion fixed at 4°C for at least 2 days.
Non-pulmonary tissue was removed from the fixed lungs and lung volume was measured via Archimedes Principle. The lungs were then embedded in 3% agar and sliced into 1.5 mm slabs and either the even or odd slabs were kept for analysis, as decided by a coin flip. Slabs were vacuum cycled, post-fixed with 1% OsO4 in 0.1 M cacodylate buffer for 2 hours, and then overnight in half saturated aqueous uranyl acetate to prevent tissue shrinkage (12, 13). The following day the slabs were vacuum cycled, dehydrated with a graded acetone series, and embedded in Technovit 7100 (Heraeus Kulzer, Wehrheim, Germany). Blocks were sectioned with a tungsten carbide knife at 1.5 μm and stained with toluidine blue.
We utilized the American Thoracic Society (ATS) recognized gold standard of stereological quantification (14) and followed a similar approach to our previous studies e.g. (9, 10, 15, 16). Slides were scanned with a 40X objective lens using an Olympus VS120 slide scanner and then systematic uniform random sampling (SURS) was performed at three digital magnifications using the Visiopharm software. A cascade sampling design was employed with 100% of the tissue sampled at 10X magnification, 30% at 20X magnification, and 10% at 40X magnification. Point counts were used to determine the volume fractions of parenchyma, non-parenchyma, ductal air, alveolar air, aerated septa, collapsed septa, and edema. Volume fractions were multiplied by reference volumes to obtain the volume of each of those compartments. Line intercepts were used at 40X to estimate surface area fractions, which were multiplied by the reference volume of parenchyma to determine the gas exchanging surface area. Mean linear intercept lengths (MLI) were estimated by multiplying 4 times the volume of parenchymal air over the septal surface area, as described in (17).
We considered five potential VILI Cost Functions to describe the amount of ongoing VILI. The dynamic driving pressure (PDDyn) was calculated for each breath from the pneumotachometer data as the PEEP subtracted from the plateau pressure, which we define as the pressure immediately prior to opening of the expiratory valve. Note that this differs from the typical clinical evaluation of driving pressure as the difference between pressure measured during an inspiratory hold and the set PEEP. As such, we refer to this waveform-derived quantity as the dynamic driving pressure. The delivered tidal volume (Vt, in ml/kg ideal body weight) was determined from measured volume, which was computed as the time integral of flow. Mechanical power was calculated by time integration of the product of pressure (P), flow (Q), and the constant 0.098 RR. This integration was performed during the inspiratory phase to compute POWI and during the entire breath to calculate POWA.
Power was also calculated according to Gattinoni as described in (18) by summing the elastic (first curly brace), resistive (second curly brace), and PEEP (third curly brace) components and multiplying by a constant to yield power in mJ/ POWG=0.098RR12Vt2Ers+RrsVt2Tinsp+Vt·PEEP. Here, Ers is the elastance according to the single compartment model for each breath, Rrs is the single-compartment resistance of the respiratory system, and Tinsp is the duration of inspiration calculated from the measured voltage at the ventilator inspiratory valve.
All cost functions were calculated on a per-breath basis from the continuous pneumotachometer recordings and the average value was used for the analysis. An average of 44,583 breaths were analyzed per mouse for a total of 2,808,710 breaths across all groups.
Linear mixed effects models were fit using the fitlme function in MATLAB 2023b with fixed effects for initial injury status (CTL or LPS) and ventilation PEEP (or unventilated). The intercept of each animal was included as a random effect. Pairwise comparisons were performed with the coefTest function and the Sidak correction for multiple comparisons was applied for relevant groups of comparisons. Significant differences between groups are indicated by * (p<0.05), ** (p<0.01), *** (p<0.001), and **** (p<0.0001).
The initial analysis is conducted using predictor values averaged over the entire four-hour ventilation period for each mouse. We then analyze the per-mouse average of 5-minute segments of predictor data sampled at 2.5-minute intervals. This rolling window approach yields 96 timepoints for regression model evaluation. Finally, we consider a per-mouse averages in a 5-minutue window centered 85 min into the ventilation period.
An elastic net regression was performed to determine which of five candidate parameters (VILI cost functions, above) was the strongest predictor of VILI as measured in changes in lung mechanics. We chose to employ the elastic net, rather than least absolute shrinkage and selection operator (LASSO) or ridge regressions, based on our desire for model interpretability and the correlations that exist between covariates (e.g. POWI and POWG). Lasso regression forces the value of some predictors to be zero. This variable selection improves interpretability but does not perform well with correlated predictors as is the case in the current model. Ridge regression avoids that pitfall and performs well with correlated predictors, but model interpretability suffers since the covariates have nonzero coefficients. Accordingly, we elected to use an elastic net with an equal weighting of ridge and lasso regression (α = 0.5). Supplemental Figure S1 shows that the coefficient values are relatively stable for 0.1 ≲ α ≲ 0.8 and the physiologic interpretation remains consistent across 0 < α ≤ 1. Prior to the regression analysis, the predictors were centered and scaled to allow interpretation of the regression coefficients without concern for the magnitude of the predictors. This is important because, for example, the magnitude PowG is about 5x greater than PDDyn.
Three elastic net regressions were performed using the lasso function in MATLAB to predict the change in elastance over the ventilation period, the change in inspiratory capacity, and the change in quasi-static compliance. In those analyses, the regularization coefficient (λ) determines the degree model simplicity. At λ = 0 there is no regularization, and each parameter is fully weighted such that the injury prediction model may be overfit. As λ gets larger the model becomes more simplified. At the largest values of λ, the prediction curve is oversimplified to a constant. A 5-fold cross validation was performed with 80% of the data used for training and 20% for testing. Supplemental Figure S2 shows that increasing CV folds from 2 to 11 reduces variability between model runs without substantially affecting the magnitude of the regression coefficients. In Supplemental Figures S1 and S2, and elsewhere in the elastic net analysis, we consider an ensemble of 1000 model runs to quantify the uncertainty of the parameter estimates. In each model run, the error computed for the testing set was used to generate mean squared error (MSE) bands for each value of λ. We report coefficient values at the values of λ where the MSE is within one standard error of the minimum MSE (λ1SE).
Dynamic driving pressure (PDDyn) was identified as the strongest predictor of VILI using the aforementioned elastic net regression. Four linear regression models were then generated for each VILI outcome measurement (change in elastance, change in inspiratory capacity, and change in quasi-static lung compliance). The first model is defined as y ~ 1 + Treatment * PDDyn and it incorporates all of the experimental data (CTL and LPS) and allows for interaction between the treatment and driving pressure terms. The second model does not include a treatment term and is defined as y ~ 1 + PDDyn. The third and fourth models are identical to the second model but only utilize CTL or LPS data, respectively. Values for R^2^, and p-values for the model, intercept, slope, treatment, and interaction terms can be found in Table S1 in the supplemental material.
Prior to mechanical ventilation, LPS-injured lungs are not significantly stiffer than controls when measured at PEEP=3 cmH2O (Elastance, Figure 2 A), nor do they have lower PV loop inspiratory capacity (IC, Figure 2 B) or quasi-static compliance (Cst, Figure 2 C). Elastance measured at PEEP=0 and 9 cmH2O was significantly increased in LPS (Supplemental Figure S3). Following four hours of mechanical ventilation, lung function was re-measured; the change in lung function is shown in the second row of Figure 2. Here, the control lungs demonstrate resiliency at each ventilation pattern and show no difference between CTL PEEP1, PEEP3, or PEEP8 groups where the ventilation-induced changes are all near zero. By contrast, LPS-injured animals exhibit PEEP-dependent changes in lung function. The PEEP1 and PEEP3 ventilation show significantly higher changes in elastance (Figure 2 D) and significantly more negative changes in inspiratory capacity (Figure 2 E) and compliance (Figure 2 F) than the PEEP8 group. The functional imparment at PEEP1 and PEEP3, but not PEEP8, yields significant differences in the change in mechanics (Figure 2 D, E, F) between CTL and LPS after PEEP1 and PEEP3, but not PEEP8, ventilation. There were no significant differences in Newtonian resistance (Rn, Supplemental Figure S4 E-H) or the tissue damping parameter (G, Supplemental Figure S5 E-H) in LPS PEEP8 mice compared to CTL PEEP8. In other words, PEEP1 and PEEP3 ventilation induced pathological changes in lung function in LPS but not CTL (Figure 2 D–F) while no differences were observed between CTL and LPS at PEEP8.
Intratracheal instillation of LPS causes inflammation-induced acute lung injury, partially mediated by the NF-κB pathway (19). Figure 3 shows the log10 transformed relative gene expression in lung tissue (RT-qPCR) and Figure 4 shows BALF protein content (ELISA) of IL-6 (A), CXCL-1 (B), and TNF-α (C). IL-6 is a sensitive pro-inflammatory cytokine, and gene expression is upregulated in all ventilated CTL groups compared to CTL NV (Figure 3A). Furthermore, CTL PEEP8 has higher IL-6 expression than CTL PEEP1. No difference was observed in CTL BALF IL-6 protein (Figure 4A). Expression of IL-6 is strongly upregulated in LPS-injured animals so that gene and protein expression of all LPS subgroups are significantly higher than the corresponding CTLs. There are no significant differences in IL-6 gene expression or protein content between any of the ventilated LPS groups.
The Log10-transformed gene expression of CXCL-1 in lung tissue (Figure 3 B) and BALF CXCL-1 protein content (Figure 4 B) is significantly higher in every LPS-injured group compared to their ventilation-matched controls. The PEEP3 ventilation group showed significantly higher gene expression than the non-ventilated animals for both CTL and LPS. As with IL-6 and CXCL1, the TNF-α gene expression (Figure 3 C) and BALF TNF-α protein levels (Figure 4 C) were significantly higher in all LPS groups than CTLs. With LPS, both PEEP1 and PEEP3 ventilation increased TNF-α gene expression compared to NV. BALF TNF-α protein content was higher in LPS PEEP3 than LPS PEEP1. In CTL, the PEEP3 ventilation increased TNF-α compared to both NV and PEEP1. As in our prior work (9, 16), NF-κB expression was not significantly affected by LPS or ventilation at this time point (Supplemental Figure S6). Taken together, the protein and relative gene expression data show that LPS induces severe inflammation prior to ventilation (the NV groups). The addition of ventilation causes further modest increases in inflammation, particularly at lower PEEPs.
Total protein concentration in the bronchoalveolar lavage fluid (BALF) was not different in any control group (Figure 3 D). Likewise, BALF albumin protein content was not significantly affected by ventilation in CTLs, although it trended higher in CTL PEEP1 (Figure 4 D). LPS injury significantly increased BALF total protein and BALF albumin for all LPS groups compared to their corresponding CTLs. The LPS PEEP8 group expressed the highest BALF total protein and the LPS PEEP3 group expressed the lowest, even compared to LPS NV. BALF albumin was significantly lower in LPS PEEP3 than LPS PEEP1.
Representative images of two whole lobes and one high magnification field from a single mouse in the CTL PEEP8 (first row) and LPS PEEP8 groups (second row) are shown in Figure 5. All lungs were perfusion fixed through the vasculature at an air inflation pressure of 5 cmH2O, after a recruitment maneuver, to capture the morphology in the physiologic breathing range (20). The CTL animals predominantly displayed healthy, homogenously inflated parenchyma. Small areas of collapse and septal folding (microatelectases), as well as larger regions of atelectasis, were identified by two or more layers of septal capillaries (Figure 5 C, arrow) stacked atop one another (Figure 5 C and F, red circle). Fully aerated alveolar septa are characterized by a single layer of septal capillaries (Figure 5 C, green circle). LPS-injured animals display markedly heterogenous collapse at both an inter- and intra-lobe level. The representative lobes shown for LPS animals (Figure 5 D, E) were harvested from the same animal. The lobe in Figure 5 D has healthy architecture and appears similar to CTL. However, the lobe in Figure 5 E displays severe collapse in the upper third, moderate collapse in the middle third, and appears relatively healthy in the lower third. Striking differences between ventilation subgroups were not observed.
The volume of parenchymal air (air in the alveoli and alveolar ducts) was mostly unchanged across all subgroups, except for a significant volume loss in LPS PEEP8 compared to CTL PEEP8 (Figure 6 A). Control lungs further exhibit resiliency with no differences in the volume of collapsed with or without ventilation (Figure 6 B). In contrast, LPS-injured lungs have significantly increased volume of collapsed septa compared to controls prior to mechanical ventilation due to the acute lung injury. The volume of collapsed septa is significantly increased with LPS PEEP1 ventilation, but not LPS PEEP3 or PEEP8, compared to LPS NV. Across all ventilation groups the LPS-injured animals display more collapsed septa than their counterpart controls due to the initial ALI.
The surface area fraction (Figure 6 C) describes the amount of gas-exchanging septal surface area per unit volume of the parenchyma. LPS animals have a significantly reduced surface area fraction compared to CTLs at both PEEP1 and PEEP3. LPS PEEP8 animals are not significantly different from CTL PEEP8 although, like all the LPS groups, they trend lower than CTL. Within the CTLs, the PEEP8 ventilation pattern yields a lower surface area fraction compared to CTL NV, suggesting modest airspace enlargement. The MLIs (Figure 6 D) offer some support for this inference with the CTL PEEP8 trending higher than the other CTLs. In the LPS groups, MLIs for PEEP1 and PEEP3 trend higher than NV and PEEP8, mirroring the trends observed in the surface area fraction. The increased LPS PEEP3 MLIs approach significance when comparing to LPS NV (p=0.087) and LPS PEEP8 (p=0.051). Additional morphometric parameters are provided in Supplemental Figure S7.
One goal of this study was to assess predictors of VILI, which we term VILI cost functions (21), in a cohort of CTL and LPS mice subjected to ventilation at PEEP = 1, 3, or 8 cmH2O. We begin by considering these VILI cost functions averaged over the entire 4-hour period of ventilation to describe the characteristics and explain the statistical model. Figure 7 shows the time-averaged candidate VILI cost functions (vertical axes) for each treatment group (horizontal axes) and ventilation cohort (colors). These variables were calculated on a per-breath basis from pneumotachometer data, and the per-mouse averages are plotted. The dynamic driving pressure (PDDyn, Figure 7 A) is calculated as the difference between PEEP and end-inspiratory (plateau) pressure and is significantly elevated at PEEP1 in LPS. The minimum driving pressure is observed at PEEP8 in LPS and PEEP3 in CTL. The tidal volume (Vt, Figure 7 B) is the delivered volume during the breath (in ml/kg) which trends slightly higher in CTL than LPS and demonstrates a modest degree of inter-subject variation. Due to this experimental variability, we elect to include this independent variable in the regression analysis. One form of mechanical power is computed by integrating the product of pressure and flow during inspiration (PowI, Figure 7 C) and increases with PEEP in both CTL and LPS, with CTL PEEP8 showing the highest values. When integrated over the entire respiratory cycle to compute energy transferred into the lungs (PowA, Figure 7 D) the LPS PEEP1 group shows the highest values, although there is not a strong separation between treatment and ventilation groups. Mechanical power is also calculated per-breath according to Gattinoni (PowG, Figure 7 E) using variables that are available from a typical clinical ventilator (18). Here, as in prior publications (18), PowG closely follows PowI. In addition, PowG is decomposed into the elastic (Figure 7 F), resistive (Figure 7 G), and PEEP components (Figure 7 H) to elucidate the contributing mechanisms. The elastic component closely follows PDDyn, the resistive component shows similar behavior to PowA, and PEEP component, which has the largest magnitude, shows similar characteristics to PowI and PowG.
An elastic net regression was performed using data from the entire 4-hour ventilation period to determine the relative importance of the candidate VILI cost functions in predicting changes in lung mechanical function. Lung mechanical function parameters are used for the response variable because the availability of both pre- and post-ventilation data allows accommodation for LPS-induced effects that are present prior to ventilation. Furthermore, those variables are in concordance with all other experimental outcomes except for BALF protein in LPS PEEP8. The predictor variables (the VILI cost functions) are centered and scaled (normalized) prior to the analysis to provide equal weighting and thus the coefficient values can be interpreted directly without consideration of the magnitude of the predictors.
Figure 8 shows the regression coefficient values (1^st^ row, A) and trace plot (2^nd^ row, B) as a function of the regularization parameter (λ) for a representative elastic net model run out of the ensemble of 1000 runs for the change in Cst. At λ = 0, on the right-hand side of the plot, there is no regularization and the model is over-fit. In other words, the coefficient values for ineffective predictors remain high (Figure 8 A). At large values of λ, shown on the left-hand side of the plot, the model is oversimplified and eventually reduces to a constant (all coefficients are zero in Figure 8 A). The green dashed line shows λMinMSE, which is the value of λ where the model mean squared error is at a minimum (Figure 8 B). The blue dashed line shows λ1SE, the value of λ located within one standard error of λMinMSE, which is used to define the ‘best’ value of λ. The standard error of the MSE is calculated from the five cross validation folds. Similar plots for the change in elastance and change in inspiratory capacity are provided in Supplemental Figure S8.
Table 1 provides the mean and standard deviation, computed with an ensemble of 1000 model runs, of the model coefficients at λ1SE. The standard deviations in Table 1 are included to show the uncertainty of the parameter estimates. Dynamic driving pressure has the largest normalized coefficient across all three of the response variables analyzed indicating that, in this data set, it is the strongest predictor of VILI-induced lung mechanical function impairment.
Given the relevance of PDDyn as a predictor of VILI, we fit linear models to predict the ventilation-induced change in elastance measured at PEEP = 3 cmH2O (Figure 9 A), the change in inspiratory capacity (Figure 9 B), and the change in quasi-static compliance (Figure 9 C). We consider a full model with interaction between PDDyn and injury (LPS or CTL), shown in Figure 9, as well as models without interaction terms that predict outcomes from PDDyn using the complete dataset, the LPS data, and the CTL data that are presented in the supplementary material. Figure 9 show predictions and confidence bands for CTL (blue) and LPS data (red) using the full model with interactions. The associated summary statistics for all models can be found in Supplemental Table S1 and show that the full model with and without an interaction term, as well as the model using only LPS perform well. In contrast, fits to just CTL are generally poor because of the lack of injury in those ventilation groups. This is also reflected in the prediction confidence intervals shown for the CTL and LPS groups, using the model with interactions, in Figure 9.
The analysis above considers the values of the VILI cost functions averaged over the entire four-hour ventilation period to elucidate the parameter characteristics, describe the statistical modeling approach, and smooth temporal variations. As such, those data show that PDDyn describes ongoing VILI, and include data immediately prior to the VILI assessments. However, in clinical application the VILI cost functions would be used to predict VILI that has yet to occur using a time-limited dataset. To test the performance of the cost functions under more relevant conditions, the elastic net modeling was repeated using 5-minute windows of the ventilator waveform data in 2.5 min steps spanning the 4-hour ventilation period with 1000 ensemble model runs conducted at each of the 96 timepoints. Figure 10 shows the temporal evolution of the model MSE (red), mean regression coefficient values (other lines), and standard deviations (bands). Overall, the coefficient values follow the same trends as the time-averaged analysis (Figure 8, Figure 9, and Table 1) with PDDyn showing the highest magnitudes and thus the strongest predictive potential. The model error (MSE, red line) generally decreases over the first hour, indicating that using 5-minute windows of data from times greater than one hour from the start of ventilation yields a more accurate prediction of changes in respiratory mechanics. There are prominent cyclic variations in the MSE and coefficient values (particularly PDDyn). These cyclic decreases in model fit quality immediately follow the recruitment maneuvers that were applied on 30-minute intervals (vertical gridlines). The elastic net behavior for elastances measured at PEEP = 0, 6, and 9 cmH2O follows similar trends and is shown in Supplemental Figure S9.
Linear models were fit using PDDyn averaged over a 5 min window that was 85 min into the ventilation period (the timepoint indicated with arrows in Figure 10). This timepoint was empirically selected to occur far enough into the experiment to avoid the transitional period at the start of ventilation, avoid the RM transients, and leave time after the PDDyn data for additional VILI to accumulate. The regressions, shown in Figure 11, are similar to the regressions fit to PDDyn averaged over the entire four-hour ventilation period, as detailed in Table 2, and the R^2^ values are a bit higher for the short time window analysis than the models using the entire dataset.
The mainstay of ARDS management is supportive mechanical ventilation with low tidal volumes in conjunction with supportive PEEP as indicated by the ARDSnet clinical trial (8, 22). However, the best approach for setting PEEP remains controversial and VILI still occurs, particularly in patients with pre-existing lung injury. One of the primary challenges in adjusting the ventilator to optimize lung protection is that there is no well-established, real-time indicator of ongoing VILI that is available at the bedside to guide adjustments. Given that VILI is caused by the forces generated by ventilation, and modern clinical ventilators can provide continuous pressure, flow, and volume measurements (waveforms), we sought to identify signals in those data that quantify VILI pathogenesis in control and lung-injured mice. Based on terminology from numerical optimization, we term these variables ‘VILI Cost Functions’. In order to test different VILI cost functions, acute lung injury was induced in mice using intratracheal endotoxin (LPS) or not (CTL). These mice were subjected to comprehensive lung function assessments, and subgroups were mechanically ventilated with PEEP = 1, 3, or 8 cmH2O. Lungs from ventilated and non-ventilated groups were harvested for morphometry and biomarker analysis to quantify the injury landscape.
In unventilated mice, IT LPS causes severe lung inflammation as evidenced by increased relative gene expression of IL-6, CXCL1, and TNF-α in tissue; increased IL-6, CXCL1, and TNF-α proteins in the BALF, and increased airspace total protein and albumin content (Figure 3 and Figure 4). LPS exposure also causes widespread and spatially heterogeneous collapse in the parenchyma (Figure 5 D–F; Figure 6 B), leading to a modest derangement in lung function. For example, elastance (stiffness) is not significantly different from controls when measured at PEEP = 3 cmH2O (Figure 2 A); however, significance arises at the extremes of the measurement spectrum (PEEP = 0 and PEEP = 9 cmH2O) (Supplemental Figure S3 A-C). Taken together, these data show that LPS profoundly impacts lung structure but has a lesser (but significant) impact on lung function, as we have reported previously in a two-hit model of LPS and high-pressure VILI (9). Note that the degree of LPS-induced function impairment depends on many factors including LPS dose, LPS origin, and instillation technique and other groups have reported more substantial alterations in lung function, e.g. (23).
The inflammatory environment and structural changes caused by IT LPS predispose the lungs to VILI that depends on the ventilation pattern (PEEP1, PEEP3, or PEEP8). In contrast, healthy lungs are resilient and no tested MV pattern affected their functional state. The LPS PEEP1 group displayed the most impaired structure and function with significant changes in stiffness, inspiratory capacity, and quasi-static compliance compared to LPS PEEP3 and LPS PEEP8, as well as corresponding CTLs. We attribute the more severe VILI at PEEP1 to LPS-induced increased atelectasis (Figure 6 B) (9, 16) leading to atelectrauma, stress concentrations, and ‘baby lungs’ (24). The deleterious effects of PEEP1 in LPS pre-injured lungs is further described in the collapsed septal tissue volume, which shows a significant increase compared non-ventilated LPS and LPS PEEP3. It is important to note that we measured structure at an air-inflation pressure of 5 cmH2O after deflation from 30 cmH2O, so the structural data corresponds roughly to the PEEP = 6 cmH2O mechanics (Supplemental Figures S3-S6, C, G). The LPS PEEP3 group fared a bit better than PEEP1 and demonstrated a significant reduction in collapsed septal tissue volume compared to LPS PEEP1. However, most lung mechanics metrics were not significantly improved at PEEP3 compared to LPS PEEP1.
The most protection, in terms of lung mechanics, for LPS was realized with PEEP8 the LPS PEEP8 lungs experienced significantly less stiffening (Figure 2 D, F) and preserved inspiratory capacity (Figure 2 E) compared to LPS PEEP1 and LPS PEEP 3 lungs. However, the quantity of collapsed lung in LPS PEEP8 morphometry was not different than LPS NV, PEEP1, or PEEP3. This is not entirely remarkable since, in previous studies, we found a weak correlation between the quantity of ventilated (open) alveoli and elastance of LPS injured lungs (9, 16). In contrast, in a model of pure VILI, elastance is strongly correlated with the number of open alveoli (15). The downward trends in surface area fraction (Figure 6 C) and upward trends in MLI (Figure 6 D) in LPS PEEP1 and PEEP3 suggest that those ventilation patterns are leading to a modest airspace enlargement and redistribution of parenchymal air. Overall, in LPS injured mice, there were not marked differences in upregulation of pro-inflammatory cytokines (Figures 3 and 4 A, B, and C). Taken together, these data indicate that, in our hands, the PEEP8 MV strategy is more protective for IT LPS pre-injured murine lungs.
Somewhat surprisingly, airspace protein content (Figure 3 D) in LPS PEEP3 was lower than in LPS PEEP8, and BALF albumin trended lower (Figure 4 D), which is not in alignment with the degree of VILI described in the mechanics or gene expression data. We speculate that the protein increase could be due to the partial pressure-induced recruitment collapsed regions that are present prior to ventilation, and the subsequent release of proteinaceous edema trapped in those atelectatic regions during ventilation (e.g., Figure 5 B). However, estimates of collapsed tissue volume (Figure 6 D) do not show evidence of additional recruitment in LPS PEEP8. Alternatively, the higher static strain in PEEP8, coupled with the pre-injured alveolar septa due to LPS, may induce alveolar leak. Another mechanism that may explain the paradoxical mechanics and total protein the LPS PEEP8 group is protection of the pulmonary surfactant from mechanical degradation by avoiding partial or complete alveolar derecruitment, and the subsequent rupture of the interfacial film on re-expansion (25), so that mechanics remain protected even in the presence of increased BALF total protein. A final possible explanation is fluid trapping during the BAL procedure, wherein BALF was not recovered from regions of locally high (or low) protein content due to, e.g., airways collapse, yielding a biased assessment of lung airspace protein content. The elevated Newtonian resistance in LPS PEEP1 and PEEP3 vs PEEP8 (Supplemental Figure S5) suggests that this may be possible.
The PEEP-responsiveness of the LPS group, where PEEP8 was significantly more protective than PEEP3 or PEEP1, stands in contrast to CTL. In those initially healthy mice, PEEP8 caused minor increases in inflammation (IL-6 expression, Figure 3A), signs of slight airspace enlargement in the surface area to volume ratio (Figure 6C), and modestly higher driving pressures (Figure 7 A). This observed lack of a strong VILI-PEEP correlation in CTL, but not LPS, aligns with clinical work including a meta-analysis showing higher PEEP was beneficial in moderate to severe, but not mild, ARDS (26). Additionally, randomized trials in non-ARDS surgical (27, 28) and ICU patients (29) have shown little to no change in outcomes with low vs. high PEEP ventilation strategies, similar to our observations in the current study.
Five parameters were selected as potential VILI cost functions to describe the amount of ongoing VILI due to their prevalence in the clinical and experimental literature, e.g. (18, 22, 30–34). We first analyzed these data averaged over the entire 4-hour course of ventilation. An elastic net regression was performed to determine the relative importance of those VILI cost functions in predicting changes in lung mechanical function. Dynamic driving pressure was found to be the most important predictor as indicated by the elastic net regression coefficient values (Figure 8 and Table 1). Tidal volume, PowI, and PowA show non-zero coefficient values for some response variables, but the confidence intervals of these parameter estimates include zero. The PowG showed a significant correlation with VILI outcomes, however this relationship was the inverse of what is expected, indicating that higher power should be lung protective.
Power during the inspiratory phase calculated by waveform integration (POWI) and using a simplified equation (POWG) were quite similar to each other as previous work in humans has shown (18). In the current study the inspiratory powers do not predict VILI as the values are highest for CTL PEEP8 followed by LPS PEEP8 (Figure 7 C and E) and those groups suffered the least VILI as reflected in lung mechanics. Decomposing the power into elastic, resistive, and PEEP components (Figure 7 F–H) reveals that the high power for the PEEP8 groups is driven by the PEEP component (Figure 7 H), which is inversely correlated to the observed degree of VILI in the current study. This stands in contrast to prior work by Rocco et al in LPS injured rats that were subjected to 1 hour of ventilation over a range of PEEPs to show that the combination of the elastic and PEEP components of mechanical power was the strongest predictor of VILI, which was predominantly measured in terms of inflammation (35). There are several potential sources for this difference such as the species, injury severity, VILI assessment (inflammation vs. lung mechanics), and duration of ventilation. From our perspective, this difference likely occurred because those rats showed the most lung protection at PEEP = 3 cmH2O and injury (inflammation) progressively increased with PEEP. In contrast, our mice suffered the least VILI at high PEEPs.
Another important consideration when comparing VILI cost functions, particularly under controlled experimental conditions, is which ventilation parameters that are varied during data generation. For example, in in Cressoni’s seminal mechanical power study in pigs (36), PEEP was held relatively constant (3–5 cmH2O), high tidal volumes of 38 ml/kg were applied, and the respiratory rate was adjusted to titrate power. Under those conditions, strong correlations were shown between the mechanical power and VILI pathogenesis due to the contributions of RR. In contrast, we fixed RR and varied PEEP. The respiratory rate is known to be an important contributor to VILI, e.g. (36–38), and is not accounted for in the driving pressure. One avenue to addresses RR in a driving pressure-like metric is through the elastic component of mechanical power (0.098 x RR x Ers x Vt^2^) which tracks closely with PDDyn (≈ Ers x Vt) when tidal volumes are similar (Figure 7 A, F). The elastic component of power linearly increases with respiratory rate while excluding the inverse correlation between PEEP and VILI found in the current study and elsewhere, e.g., (39–41). The predictive performance of the elastic component of mechanical power has been demonstrated in other work (42).
Linear regressions for each driving pressure-VILI outcome relationship (Figure 9, Table S1) show that dynamic driving pressure is a weak predicter of lung function impairment when applied only to CTL animals, which is to be expected because CTL animals have resilient lungs and are not functionally affected by the VILI challenges. However, driving pressure is a strong predictor when applied only to the LPS data or to pooled LPS and CTL data, where higher driving pressure predicted an positive change in elastance (Figure 9 A), a negative change in inspiratory capacity (Figure 9 B), and a negative change in quasi-static lung compliance (Figure 9 C), each with strong values for R^2^ and highly significant p-values for the model (Supplemental Table S1). Because use of the dynamic driving pressure to guide ventilation would employ regression equations fit to a continuum of injury severity these models support the use of driving pressure for ventilation adjustment independent of injury status, at least in healthy or endotoxin-injured mice. Our regression analysis parallels a study by Samary et al in rats injured with endotoxin (LPS) and ventilated for one hour with different combinations of tidal volume and PEEP to show positive correlations between driving pressure and measures of inflammation (32). Note that the mechanical power assessment by Rocco et al (35) discussed above is a re-analysis of the Samary et al data. The alignment of our findings in mice with Rocco’s driving pressure-based analysis, but not the power-based analysis, suggest that PDDyn may be a more robust predictor.
The likely clinical use case for PDDyn is to analyze short segments of data to determine if a ventilation strategy is optimally lung protective. In the analyses discussed above, we considered the entire 4-hour span of data to smooth out temporal variations and simplify presentation of the somewhat complex statistical analysis. To evaluate the potential for predictive use of PDDyn, we applied the elastic net regression analysis to 96 5-minute windows of PDDyn spanning the 4-hour ventilation period (Figure 10). Those data indicate that even short periods of ventilator waveform data can be used to predict future VILI. Figure 10 also shows that there is an initial ‘run in’ period, spanning approximately one hour, during which the predictive quality of the model (e.g. the MSE) improves. Importantly, this data, and the linear regressions using data from 85 min into the ventilation Figure 11, show that PDDyn can be used to predict changes in lung mechanics occurring 2.5 hours later. Furthermore, there are cyclic decreases in model quality, and corresponding changes in coefficient estimates, which correspond to the application of recruitment maneuvers. The model predictive accuracy generally tracks with the magnitude of the inter-group PDDyn differences (Supplemental Figure S10) which are smallest immediately following an RM and diverge more substantially over the course of the ventilation period.
A number of observational and retrospective clinical studies have shown that decreased driving pressure is associated with increased survival in ARDS patients, e.g. (43–46). Those studies generally show similar results to Amato et al’s large secondary multilevel mediation analysis of 3,562 ARDS patients from nine randomized trials that found driving pressure was most strongly associated with survival, as opposed to tidal volume or plateau pressure (47). However, the use of driving pressure to prospectively set PEEP in ARDS patients has yet to show improved outcomes, although safety and feasibility have been shown in smaller studies, e.g. (48–50).
One limitation of the current study is that we only tested only female mice, so sex differences were not considered and this is certainly an area for future study. Inflammatory lung injury was generated with intratracheal endotoxin, but ARDS can arise from many types of lung injury and future work should be conducted to assess cost function predictive power in additional models of lung injury. Finally, the duration of mechanical ventilation was four hours which is significantly less than the typical course of supportive care for an ARDS patient.
In summary, the goal of this study was to test the efficacy of parameters to guide real-time ventilator adjustments to reduce VILI. Mice with and without LPS-induced acute lung injury were ventilated for four hours at different PEEPs and the effects of VILI were measured in lung structure, function, and inflammation. The data show that PEEP = 1 cmH2O ventilation of LPS-injured mice was most injurious. Regression analysis reveals that dynamic driving pressure (PDDyn), but not tidal volume or forms of mechanical power, was a strong predictor of VILI as quantified by changes in lung mechanical function. The predictive power of PDDyn improved after a 1-hour ‘run in’ period and was worse following recruitment maneuvers. These findings support the concept of utilizing driving pressure to determine optimally lung-protective ventilation settings.
Supplemental Data
Supplemental Table S1 and Supplemental Figures S1-S10: https://doi.org/10.6084/m9.figshare.27245307