David W Kaczka

Accurate and rapid characterization of lung mechanics remains a central challenge in respiratory disease management. Physics-informed poroelastic finite-element (FE) models resolve detailed tissue–airflow interactions but are computationally prohibitive for real-time or large-scale clinical applications, while lumped-parameter models sacrifice mechanistic fidelity for efficiency. In this work, we present a porcine-specific, multi-fidelity computational framework that integrates poroelastic FE modeling with machine learning to enable rapid, uncertainty-aware estimation of respiratory compliance (
C
rs
) and resistance (
R
rs
). High- and low-fidelity simulations are generated from CT-derived porcine lung geometries by sampling a physiologically relevant parameter space, and the resulting pressure–volume dynamics are used in an inverse modeling procedure to infer global respiratory mechanics. A key result is that multi-fidelity Gaussian process (MF-GP) surrogates achieve accurate predictions of
C
rs
and
R
rs
with errors below 5% relative to high-fidelity simulations, while providing computational speedups of over five orders of magnitude. In contrast, neural network (NN) surrogates exhibit relatively poor generalization in the data-scarce regime considered, highlighting the importance of model selection for scientific machine learning under limited high-fidelity data availability. Beyond predictive performance, global sensitivity analysis reveals a clear mechanistic separation in parameter influence: compliance is primarily governed by elastic stiffness and chest-wall coupling, whereas resistance is dominated by permeability. The weak interaction effects observed support an approximately additive response structure, enabling robust parameter identifiability and reduced-order representations of the inverse problem. The framework is validated against independent ventilator measurements from porcine lungs, showing strong agreement within clinically observed ranges. Overall, this study provides new insight into the structure of the inverse problem in poroelastic lung modeling and establishes a computationally efficient pathway for uncertainty-aware prediction and parameter estimation, with potential applications in personalized ventilation and preclinical study design.

Introduction Characterizing the impact of distending pressure on airway structure across different generations is useful for guiding fluid-structure interaction in computational modeling of airflow in the lungs. We evaluated the impact of airway pressure and lung volume on airway lengths, diameters, branching angles, and curvature of porcine airway segments in-vivo. Specifically, we quantified pressure-induced structural changes in porcine airway segments, by analyzing CT-derived morphometry from the trachea to the fourth generation during static breath holds at multiple distending pressures. Methods Whole-lung CT scans from six porcine subjects (39.0 to 47.7 kg) under anesthesia were obtained at constant airway pressures from 30 to 0 cmH2O in decrements of 5 cmH2O. 3D Slicer and Vascular Modeling Toolkit were used to semi-automatically segment and skeletonize the airway trees from the CT images. For each subject, the same airway segments were matched across all distending pressures. Segments included the trachea, and one airway segment from the 1st to the 4th generation, for left and right lungs. Airway lengths, diameters, branching angles, and curvatures were normalized to the maximum values and averaged across all subjects. ANOVA was used to compare dimensions across pressures, with p < 0.01 considered statistically significant. Results Figure 1 shows the trend across all geometric parameters. Tracheal, as well as left and right lung diameters across all generations from 30 to 15 cmH2O were significantly different than those at 5 and 0 cmH2O. Significant differences in curvature were observed in the first generation of the right lung only. Curvatures at 25 and 15 cmH2O were significantly different from the curvature at 0 cmH2O. For branching angles, significant differences were observed in the second and fourth generations of the left lung. Branching angles for pressures between 30 and 25 cmH2O were significantly different from the angle at 0 cmH2O for both generations. No significant differences in curvatures were observed in the right lung across generations. Conclusions Our data indicate that airway diameter is significantly dependent on distending pressure for all generations examined. Tracheal diameter, as well as right and left lung airway diameters, exhibited a curvilinear dependence on pressure. Tracheal length, as well as right and left lung airway lengths, were fairly constant, while curvature and branching angles showed no consistency. These results highlight the interplay between airway mechanics and distending pressure, which may be important for improving computational airflow simulations and refining models of airway mechanics under healthy and diseased conditions. This abstract is funded by: T32 HL144461- 03, W81XWH-16-1-0434, W81XWH-21-1-0507, W911NF-23-1-0004

Rationale Numerical simulation of ventilation distribution provides quantitative insights into air flow distribution within the complex geometry of the lung. Such simulations reveal regional differences in ventilation, assess the impact of disease, and offer a noninvasive, reproducible means to study pulmonary mechanics with potential to improve respiratory care. Registering and mapping regional lung elastances derived from CT-based Jacobian determinants onto the acini of a virtual airway tree provides spatially heterogeneous properties that simulate in vivo lung mechanics more accurately. Compared to those using uniform or Perlin-based acinar elastances, the resulting simulations may yield more physiologically realistic ventilation distribution, and improve the fidelity of functional lung modeling. Method A one-dimensional (1D) airway tree of a 40 kg porcine lung was generated within a three-dimensional (3D) domain using a space-filling algorithm, yielding 30,959 airway segments and 15,479 terminal points for acinar placement. An anesthetized pig at the same weight received volumetric 3D CT scans during static breath holds at various pressure settings. The voxel-wise elastances derived from the CT images were mapped to the acini of the numerical airway tree via non-rigid registration and Voronoi partitioning. The nodes of the numerical airway tree represent the effects of viscous dissipation and convective acceleration of gas flow through cylindrical airway segments, along with the viscoelastic behavior of airway walls and surrounding parenchyma. Regional ventilation to individual acini was evaluated by traversing the entire tree with a recursive flow-division algorithm, enabling computation of acinar flow and pressure distributions. The simulations were performed across a wide range of ventilation frequencies, to examine the resulting ventilation distributions. Results Compared with the uniform acinar elastance setup, the simulations using mapped acinar properties exhibited greater regional variations in pressure (Figure 1) and flow magnitude, as well as phase difference, across the wide range of ventilation frequencies. Acinar flow magnitudes showed a substantially broader range below the resonant or corner frequency, while acinar pressure magnitudes remained relatively similar in this region. Overall, simulations incorporating mapped acinar elastances are considered more physiologically realistic. Conclusion Numerical simulation of ventilation distribution offers a noninvasive way to study lung mechanics and optimize respiratory care. By mapping CT-derived regional elastances onto acini, simulations capture in vivo spatial heterogeneity and lung realism more accurately, yielding more physiologically meaningful ventilation patterns than those using uniform or Perlin-based acinar elastances. This abstract is funded by: W81XWH-21-1-0507, W911NF-23-1-0004

Introduction: Acute respiratory distress syndrome (ARDS) is characterized by significant heterogeneity in lung mechanics, leading to ventilation-perfusion mismatch and potential for ventilator-induced lung injury. Selective lobar ventilation offers a tailored approach for managing the impact of such heterogeneity by isolating and ventilating specific lung regions. However, its clinical feasibility, particularly the time needed for intubation, has not been established. This study evaluated the time required for selective lobar intubation and ventilation using a novel airway device in a high-fidelity mannequin. The primary hypothesis was that procedural times would follow a log-normal probability distribution. A secondary hypothesis was that the probability of the procedure exceeding five minutes would be less than 5% at the 95% upper confidence limit, if a log-normal distribution were suitable.

Methods: This was a prospective observational study conducted at a university simulation center. Clinicians were invited to perform endotracheal intubation, followed by selective endobronchial intubation of the left lower lobe on an AirSim Advance Bronchi X mannequin (TruCorp Ltd., Lurgan, UK). Participants used a "tube-thru-tube" technique with video laryngoscopy and fiberoptic bronchoscopy, guided by stepwise visual instructions. Data collected included the total time from tube insertion to successful selective lobar ventilation and the participant's self-reported number of intubations performed in the previous year. The fit of procedural times to log-normal and other distributions was assessed using the Shapiro-Wilk and Pearson chi-square tests.

Results: Among all 52 participants, there was a poor fit to a log-normal distribution (P = 0.0040). However, the six participants who performed less than five endotracheal intubations in the preceding year had significantly longer times (P = 0.0006). Among the other 46 clinicians, procedural times showed a strong fit to a log-normal distribution (Shapiro-Wilk W = 0.98, P = 0.73), superior to normal or Weibull distributions. The mean time for successful selective lobar intubation was 2.26 minutes. No participant exceeded a five-minute threshold. Utilizing the log-normal model, the calculated 95% upper confidence limit on the probability of exceeding five minutes was 0.02%.

Conclusions: This simulation shows that procedure times for selective lobar ventilation follow a log-normal distribution. This statistical predictability is essential for quantitatively evaluating safety and for designing future clinical trials, including novel ARDS therapies with selective lobar ventilation. The confirmation of the log-normal distribution for an advanced airway task can be applied to other assessments of intubation times to make quantitative comparisons (e.g., ratios of means) and to calculate probabilities of exceeding tolerance limits.

Introduction: Protective Conventional Mechanical Ventilation (CMV) has substantially improved outcomes in patients with ARDS over the last two decades, although mortality remains unacceptably high. We hypothesized that multi-frequency ventilation (MFV), in which small volume oscillations at multiple frequencies are added to a CMV waveform, would improve lung aeration and ventilation homogeneity, as assessed by quantitative Computed Tomography (qCT).

Methods: Twenty-five pigs were mechanically ventilated using a hybrid ventilator/oscillator (OscillaVent Inc., Iowa City, Iowa). Lung injury was induced by intravenous infusion of oleic acid, after which the animals were randomized to receive: 1) Protective CMV; 2) high frequency oscillatory ventilation (HFOV) per established protocol; or 3) MFV with a waveform consisting of 3.5 and 7 Hz oscillations superimposed on a CMV waveform. For each group, seven whole-lung CT scans were obtained during static breath holds from 30 to 0 cmH2O, in decrements of 5 cmH2O. Images were obtained at five time points, including before and immediately after lung injury, as well as 3-hour intervals thereafter. Quantitative CT analyses, including aeration and texture assessment, were performed on each segmented image.

Results: Air volume decreased with reduced pressure at each time point and was significantly higher in HFOV compared to CMV. Aeration analysis revealed a higher percentage of non-aerated regions during CMV compared to MFV and HFOV, with significant increases with decreasing pressure. Normally aerated regions were more prevalent during HFOV than in CMV, and decreased with decreasing airway pressure. Texture analysis showed a significant increase in consolidated areas with decreasing pressure, with higher values in CMV compared to MFV and HFOV.

Conclusions: This study shows that in a porcine model of ARDS, both MFV and HFOV improved lung aeration and reduced consolidation compared to CMV. MFV demonstrated results similar to HFOV, but with lower mean airway pressures. These findings suggest that MFV may offer similar benefits in lung recruitment and ventilation homogeneity compared to HFOV, with less risk of hemodynamic impairment. Further research is needed to assess the clinical applicability of MFV in patients, and its long-term effects in ARDS management.

Introduction: Ventilatory support in ARDS typically relies on lung-protective strategies, aimed at minimizing risk for ventilator-induced lung injury (VILI). In this context, the ability to anticipate changes in respiratory function may support individualized treatment, thus improving patient outcomes. Ventilator waveforms such as airway flow, pressure, and volume are continuously monitored and can be analyzed with machine learning techniques to identify patterns associated with key physiological derangements. In this study, we investigate the use of a convolutional neural network (CNN) to estimate respiratory system compliance. Our objective is to assess whether this approach can predict changes in compliance over time in a large animal model of ARDS.

Methods: Acute lung injury was induced via oleic acid infusion into the pulmonary artery. Following injury maturation, nine pigs were ventilated with volume-controlled, lung-protective conventional mechanical ventilation and monitored for nine hours. Airway pressure and flow waveforms were recorded at five timepoints: baseline (BL), immediately after injury maturation (T0), and 3, 6, and 9 hours post-injury (T1, T2, T3). Respiratory system mechanics parameters were estimated using multiple linear regression based on the equation of motion. A CNN was developed to predict respiratory system compliance at T1, T2, and T3.

Results: Model predictions of future compliance (i.e., 3 hours later) showed a strong alignment with observed compliances at all time points. At T1, predicted compliance was slightly higher than true compliance, with no significant difference (p = 0.471), achieving the highest correlation and lowest RMSE. At T2, predicted closely matched true compliance, with strong correlation. At T3, predicted matched true compliance, with significant correlation. (r2 = 0.54, p = 0.025) and RMSE 2.89 mL cmH2O-1.

Conclusions: Overall, our findings support the capability of a CNN to capture accurately individual compliance trajectories over time. Future research should aim to validate this approach in larger and more diverse populations, and to explore the practical use of ventilator waveform data for real-time patient monitoring at the bedside. Such developments could provide clinicians with valuable tools to support clinical decision-making.

BackgroundDespite the implementation of protective mechanical ventilation, ventilator-induced lung injury remains a significant driver of ARDS-associated morbidity and mortality. Mechanical ventilation must be personalized and adaptive for the patient and evolving disease course to achieve sustained improvements in patient outcomes. In this study, we modified a military-grade transport ventilator to deliver the airway pressure release ventilation (APRV) modality. We developed a computationally-directed (CD) method of adjusting the expiratory duration (TLow) during APRV using physiologic feedback to reduce alveolar derecruitment and tested this modality in a porcine model of moderate-to-severe ARDS.MethodsFemale Yorkshire-cross pigs (n = 27) were ventilated using a ZOLL EMV+® 731 Series ventilator during general anesthesia and subjected to a heterogeneous Tween lung injury followed by injurious mechanical ventilation. Animals were subsequently ventilated for 6 hours under general anesthesia after randomization to one of three groups: VT6 (n = 9) with a tidal volume (VT) of 6 mL/kg and stepwise adjustments in PEEP and FiO2; VT10 (n = 9) with VT of 10 mL/kg and PEEP of 5 cmH2O; CD-APRV group (n = 9) with computationally-directed adjustments in TLow based on a nonlinear equation of motion to describe respiratory mechanics. Results are reported as median [interquartile range].ResultsAll groups developed moderate-to-severe ARDS and had similar recovery in lung injury, with all demonstrating final PaO2:FiO2 > 300 mmHg (VT6: 415.5 [383.0–443.4], VT10: 353.3 [297.3–397.7], CD-APRV: 316.6 [269.8–362.4]; p = 0.12). PaCO2 was significantly higher in the VT6 group compared with the CD-APRV group (59.3 [52.3–60.1] mmHg vs. 38.5 [32.7–52.2] mmHg, p = 0.04) but not significantly different from the VT10 group (47.5 [45.3–54.4] mmHg; p = 0.32 vs. VT6) despite having a significantly higher respiratory rate (30.0 [30.0–32.0] breaths/min) compared with VT10 (12.0 [12.0–15.0] breaths/min, p = 0.001) and CD-APRV (14.0 [14.0–14.0] breaths/min, p < 0.001) groups at the study end.ConclusionWe successfully implemented a computationally directed APRV modality on a transport ventilator, adjusting TLow based on respiratory mechanics. This study demonstrated that CD-APRV can be safely used, with the advantage of guiding expiratory duration adjustments based on physiologic feedback from the lungs.

Patients with acute respiratory failure often require supportive mechanical ventilation to maintain adequate gas exchange. Recent studies have shown that multi-frequency ventilation (MFV), the technique of presenting multiple simultaneous frequencies in flow or pressure at the airway opening, may provide more uniform ventilation distribution and parenchymal strain throughout the mechanically heterogeneous lung. In this study, we simulated gas flow within a porcine central airway tree, from the trachea to the fifth generation, with dynamic boundary conditions during volume-controlled conventional mechanical ventilation (CMV) cycled at 0.27 Hz (16.2 min-1), as well as MFV waveforms comprised of two fast sinusoidal components (i.e., 3.5 Hz and 7.0 Hz) superimposed on the 0.27 Hz CMV waveform. By using forced gas flows at the airway opening of the computational lung model, dynamic pressures at various airway segments were predicted, based on the interactions of internal flow with the downstream elastances and peripheral airway resistances. Internal airflows were simulated and analyzed in both time- and frequency-domains. The results indicate that MFV resulted in stronger asymmetric flow (i.e., “pendelluft”) at end-inspiration and end-expiration. MFV also appeared to augment inlet-outlet phase differences for both pressure and flow compared with CMV, suggesting that MFV may enhance gas mixing, thus facilitating more efficient ventilation.

Accurate modeling of lung parenchymal biomechanics is critical for understanding respiratory function and improving diagnoses. Traditional hyperelastic models capture tissue deformation but miss essential physiological interactions. This study evaluates an experimentally informed poroelastic model (Birzle's formulation) against hyperelastic-only models within a finite element framework. Using porcine lung geometry and CT-based boundary conditions, we simulate realistic breathing cycles and compare deformation, stress, strain, and volume change. Results show that poroelasticity better reproduces pressure-volume behavior and ventilation distribution, underscoring the importance of fluid-influenced mechanics for robust, clinically relevant lung modeling.