The diaphragm is the major muscle responsible for inspiration, and contributes to up to 70% of the inspiratory tidal volume in a healthy individual. 1,2 Diaphragm dysfunction can result from a variety of etiologies including phrenic nerve trauma 3 and neuromuscular disease 4,5 . Owing to the degenerative nature of many of these etiologies, mechanical respiratory failure exists as a continuous spectrum of dysfunction. Severe diaphragm dysfunction or paralysis can lead to chronic respiratory failure. When disease progresses beyond the treatment capacity of noninvasive treatment, patients must make the difficult decision to opt for permanent invasive ventilation via a tracheostomy or to pursue palliative care with an understanding of the terminal nature of their disease. Invasive ventilation can interfere with many aspects of a patient's quality of life, such as hindering speech, requiring full-time care, and possibly necessitating the patient move into a care facility. There is an urgent need for new therapeutic ventilation options that restore respiratory performance without sacrificing quality of life, especially for those with the most severe cases of diaphragm dysfunction.
Respiration is a fundamentally mechanical process. The diaphragm is a dome-shaped muscle that drives up to 70% of respiration 1,6 . Soft robotic actuators are ideal for reproducing complex, repetitive muscle contractions, such as that of the diaphragm, while interfacing nondestructively with biological tissue. Previously, fully implanted soft actuators have shown the ability to augment heart function [7][8][9][10][11] and many other newly developed implantable robotics have shown utility in a broad spread of biological applications [12][13][14][15][16][17][18][19][20] . Due to the mechanical nature of respiratory failure, especially in the context of conditions like muscular dystrophy, implanted soft robotic actuators applied to the diaphragm have the potential to mechanically support and augment its function. There is minimal prior work investigating soft robotics applied to the augmentation of respiration; one of the few examples reports a dielectric elastomer sheet used to completely replace an excised diaphragm and generate motion. 12,21 Contrastingly, the work presented here leaves the native diaphragm intact while demonstrating function in terms of augmentation of clinically-relevant physiological metrics (ventilation flows, volumes, and pressures) in addition to diaphragm motion in an in vivo porcine model as opposed to solely replicating diaphragm motion while excising the native diaphragm.
Here, we demonstrate a diaphragm assist system that functions as an implantable ventilator by using soft robotic actuators to mechanically augment diaphragm function during inhalation, increasing inspiration. As a proof-of-concept, we simulate a range of respiratory insufficiency within each animal-specifically, we induce respiratory depression via anesthetics and diaphragm paralysis by severing the phrenic nerve-and then demonstrate the ability of the assist system to augment respiratory flows, volumes, and pressures. We also investigate specific metrics of inspiratory function including peak inspiratory flow and transdiaphragmatic pressure 22 . We show that in order to achieve effective inspiration assistance, the actuation of the assist system must be synchronized to the subject's underlying respiratory effort. To achieve this, we have built a control system in which actuation is triggered by the beginning of inspiration. Through an analysis of the respiratory waveforms, we investigate the optimal alignment of actuation with the subject's native respiratory effort. By augmenting diaphragm function in a biomimetic fashion, we demonstrate the replication and augmentation of the native biomechanics of respiration in which a negative pleural and alveolar pressure drives airflow, as opposed to the positive pressure ventilation of standard mechanical ventilation.
As depicted in the schematic in Fig. 1a, when the diaphragm contracts, the arclength of the diaphragm shortens, and the entire sheet of the diaphragm moves downwards, acting as a pump. The thoracic cavity volume increases and pressure decreases, ultimately driving respiration.
Our strategy aims to harness the contractile function of pneumatic artificial muscles (PAMs) to mimic and augment the native contraction of the diaphragm. We opt for a McKibben type PAM-a classical soft actuator type with a simple fabrication process and high force generation 23,24 that is capable of mimicking and augmenting biological systems. 7,8,13 At their simplest, McKibben actuators are composed of an expandable weaved mesh surrounding a bladder connected to an airline (Fig. 1b) (see Methods). When the bladder is pressurized, the mesh expands radially and drives linear contraction (Fig. 1c). The McKibben actuators used in this work were capable of generating up to 40 N of contractile force under 20 psi pressurization (see Extended Data Fig. 1, and Extended Data Fig. 2, and Supplementary Notes). Conceptually, we harness the linear contraction of these PAMs by placing them superior to the native diaphragm so that the relaxed PAM conforms to the native curvature of the diaphragm (Fig. 1d). Mimicking the native diaphragm, we anchor the ends of the PAMs to the ribs (see Methods). With pressurization, the length of the PAM shortens, the arclength shortens, and the PAM mechanically pushes the diaphragm downwards (shown in situ in Fig. S1). Actuator behavior is governed by the degree of pressurization. Set pressurization waveforms are programmed to the control system and electropneumatic regulators. In vitro and in vivo characterization of actuator behavior when controlled by different pressurization waveforms is included in the Extended Data (Extended Data Fig. 1 and Extended Data Fig. 2).
In contrast to the dielectric artificial diaphragm described by Bashkin, et al., 12 our diaphragm assist system uses a set of two linear PAMs, leaves the native diaphragm intact, and has a low profile presence (deflated: 5 mL volume, inflated: 17 mL volume). To test this concept in a live porcine model, we surgically implanted a pair of McKibben actuators in an anterior-to-posterior direction lateral to the heart. The actuator placement is visualized in a 3D rendering in Fig. 1e. Fluoroscopy of the diaphragm was taken throughout the experiments. The lateral cross-sectional view from the fluoroscopy shows the realization of our soft robotic strategy in an in vivo pig model (Fig. 1f,g).
The actuators push the diaphragm caudally, augmenting the diaphragm displacement. Ultrasonography is used to visualize and quantify diaphragm displacement (Fig. 2). The coronal plane cross section of the actuator and the diaphragm is visualized via two-dimensional (B-mode, brightness) ultrasonography of the diaphragm (Fig. 2a,b). To quantify the motion of the device and the diaphragm, we used the (M-mode, motion) ultrasonography (Fig. 2c,d), which visualizes the image along a single line -selected within the B-mode image -over time. M-mode features excellent axial and temporal resolutions, and is particularly well-suited for motion analysis 25 . The actuator augments the diaphragm displacement per breath from 0.37 cm displacement of unassisted ventilation (Fig. 2c) to 1.92cm displacement of assisted ventilation (Fig. 2d). (20 psi). For all images, the probe was positioned in the right subcostal space, pointing toward the cranial direction. Orange dashed line: diaphragm, blue dashed ellipse: actuator cross-section, blue star: spatial position of the diaphragm.
To evaluate the ability of our diaphragm assist system to augment respiratory function, the animals were instrumented to collect physiological data, including respiratory flows, volumes, and pressures within the respiratory system (Fig. S2). The pressurization of the soft robotic actuators was controlled via a custom-built control system; the actuation pressure data was input into the same high-resolution data acquisition system as the physiological data (see Methods).
Ventilation is key to driving CO2 exchange, so we first examine the flow and volume waveforms as metrics of ventilatory function. Flow is measured by a spirometer. Peak inspiratory flow can be used as a clinical metric of inspiratory function 22 , which yields a direct measurement of the effect of the diaphragm assist system. Integrating the flow with respect to time yields a volume waveform over time. The volume of each breath (tidal volume) and its rate (minute ventilation) are the most relevant parameters of directly measuring ventilation. Pressures within the respiratory system, such as pleural and abdominal pressures, reveal information about the respiratory biomechanics that physically drive ventilation and are discussed later in this work.
To start each study, the animal was anesthetized appropriately with isoflurane and placed on mechanical ventilation. Isoflurane induces a respiratory depression with decreased tidal volumes and increased respiratory rate that ultimately combine to a reduced minute ventilation 26 . The respiratory depression secondary to the isoflurane is used as our baseline animal model of respiratory insufficiency due to hypoventilation. Each subject has a reduced but non-zero respiratory drive and response to CO2. Mechanical ventilation is used to support the animal throughout the implantation surgery. Within each subject, we introduce a series of respiratory challenges, collecting data during periods of unassisted ventilation (in which any spontaneous respiration is due to the native respiratory drive) and during periods of actuator assisted ventilation. Mechanical ventilation is used to restore and maintain a state of normoventilation after and between respiratory challenges. To investigate the effect of the diaphragm assist system, a representative respiratory challenge was chosen per subject. The phrenic nerve is intact for all data shown in Fig. 3.
In a vignette from the best responding subject (Fig. 3a), we show that the assist system has the direct capacity to augment the peak inspiratory flow from 0.18 L/s to 0.59 L/s and the tidal volume from 55 mL to 161 mL. When the assist is resumed after a short period of unassisted respiration, the augmentation effect of the actuation on the flow and volume waveforms is reestablished nearly immediately over the course of 2 breaths.
An example of a full respiratory challenge is shown in Fig. 3b. During the unassisted ventilation at the start of the challenge, the subject models a state of hypoventilation. During this period, the tidal volumes and flows have a slight increase over time, indicating the baseline respiratory drive is responding to the increasing CO2 status due to the unassisted low minute ventilation (0.9 L/min). When assist is switched on (as indicated by the actuator pressure waveform, the white background, and the black arrow), there is a clear jump in the peak inspiratory flow (+0.20 L/s, 95%CI +0.19 L/s to +0.22 L/s), tidal volumes (63mL, 95% CI 58 mL to 68 mL), and minute ventilation (0.9 L/min to 3.1 L/min). The actuators cycle between a pressurized and unpressurized state for 10 minutes. At the end of the respiratory challenge when the respiratory effort has reached a steady state, the assist is switched off and we see that the respiratory effort drops slightly (peak inspiratory flow: -0.09 L/s, 95% CI -0.08 to -0.10; tidal volume: -10 mL, 95% CI -7 to -13) but much less than the jump seen at the start of the respiratory challenge.
The respiratory drive is a slow but dynamic factor underlying all of the respiratory physiology data. As seen in the first 200 s of Fig. 3b, the respiratory drive visibly increases as the low minute ventilation leads to CO2 buildup. This response to CO2 is dynamic and varies between subjects based on each animal's response to isoflurane. By examining the breaths immediately before and after these transition points (off-to-on and on-to-off), we can examine the direct effect of the diaphragm assist system in terms of augmenting volume and peak inspiratory flow while minimizing the influence of the changing baseline. This analysis was conducted for one representative respiratory challenge per each of 5 subjects. These 5 subjects represent the subset of experiments conducted with uninterrupted respiratory challenges that collected the unassisted baseline data both before and after device assistance. We see a spectrum of responsiveness to the diaphragm assist system across subjects (Fig. 3c,d,e). The subjects are ordered from largest change in tidal volume at the start of the challenge to the smallest (best responder to worst responder according to Fig. 3d). We find the diaphragm assist system generates much larger respiratory augmentations at the beginning of a trial-when mechanical ventilation support has just been removed, minute ventilation drops suddenly, and the animal's CO2 state rises rapidly-than at the end of the respiratory challenge when the respiratory baseline is relatively more stabilized (Fig. 3c,d,e).
Subject A was much more responsive to the assist system than any other subject. In terms of tidal volume, 4 of the 5 subjects show an augmentation of >30 mL per breath at the beginning, whereas only 1 of the subjects shows substantial augmentation to the tidal volume at the end. Of the 4 less responsive subjects (B,C,D,E), 3 of them show a mild response at the end while in the worst responder (E), the actuation overall decreased the ventilation metrics (Fig. 3c-e). The subject with the weakest response had the highest baseline weight-normalized minute ventilation at the beginning of the trial (Fig. 3e) compared to other subjects. Body weight normalized minute ventilation is used to compare these results to normal physiology. Minute ventilation is a metric of the ventilation rate, taking into account both tidal volume and the respiratory rate. In a normal, conscious pig, the expected body weight normalized minute ventilation is 198 mL/min/kg 41 mL/min/kg with a range of 104 mL/min/kg to 262 mL/min/kg 27 , indicated by the green shading in Fig. 3e. Actuator assisted ventilation allowed all 5 subjects to reach the lower range of normal physiology, and 2 of the subjects even achieved a minute ventilation corresponding to one standard deviation below the normal mean (Fig. 3e). However, we note that this minute ventilation is achieved with low tidal volumes and high respiratory rates, which results in a lower alveolar ventilation than the same minute ventilation achieved with high tidal volumes and low respiratory rates. Synchronizing with the underlying respiratory effort. Like with standard mechanical ventilation 28,29 , patient-ventilator synchrony in our system is critical to the ability to augment respiration. Asynchronous ventilation can destructively interfere with the underlying respiratory effort, leading to worse ventilation with assistance than without.
In order to synchronize the actuation of our assist system with the subject's underlying respiratory effort, we built a control system (Fig 4a,b) that can actuate based on the respiratory flow rate. The system uses the spirometry flow sensor as the source data. The flow data is read into our data acquisition system. The associated data analysis software allows a user-set threshold voltage; this threshold voltage was manually titrated during every respiratory trial to achieve qualitatively good synchronization. When the flow rate passes this set threshold, a digital pulse is triggered and sent to the microcontroller in our control box. The microcontroller triggers a pre-set actuation pressure waveform of one cycle of pressurization and depressurization in the electropneumatic regulator, filling and emptying the PAMs with pressurized air (further details in Methods).
Our control system can implement both a set, rhythmic control scheme independent of the native respiratory effort or a dynamic control scheme synchronized with the underlying respiratory effort. Due to the phase and frequency mismatch between the independent actuation and the underlying respiratory effort, the mixed interference of the actuator and the underlying respiratory effort can be seen in both the flow and volume waveform (Fig. 4c). Contrastingly, the well synchronized actuation reveals much more homogenous flow and volume waveforms. (Fig. 4d).
Within each subject, we compare the tidal volumes and peak inspiratory flows in one representative challenge of independent actuation with one representative challenge of synchronized actuation (details in methods). We find that synchronized actuation consistently produces much less variance in the tidal volumes (Fig 4e,f). Although in some subjects-such as subject A-independent actuation achieved a few higher maximum tidal volumes, the independent actuation also achieved lower minimum tidal volumes across all subjects due to the misalignment of actuations with the underlying respiratory effort leading to destructive interference or due to actuation with no underlying breathrepresenting a breath that is solely actuator driven. Misalignment between the diaphragmatic contraction and the device during independent actuation can be observed with M-mode ultrasound (Fig 4g), in contrast to synchronized actuation (Fig 4h). Asynchronous moments of native diaphragm contraction produce to a heterogenous waveform, and indicated by the orange arrows in Fig. 4g.
Physiologically, ventilation is necessary to bring in oxygen (O2) and to clear out accumulated carbon dioxide (CO2) from the blood. Arterial blood gases (ABGs) are discrete blood analyses that give a snapshot view of the gas exchange and acid-base homeostasis, providing measurement of partial pressure of O2 (PaO2) and CO2 (PaCO2), pH, and bicarbonates (HCO3 -) in arterial blood. PaCO2 is directly and inversely proportional to alveolar ventilation and is therefore a representative metric of ventilatory function. Only pH and pCO2 are depicted here in Fig. 5, but the full ABG parameters are reported in Extended Data Table 1 and discussed in the Supplementary Notes.
As shown in the prior section, the high variance from independently actuated ventilation showed mixed constructive and destructive interference (Fig. 4e,f) which led to worse ventilation outcomes. The same variance in the peak inspiratory flows and tidal volumes over time due to independent vs. synchronized actuation can be seen in Fig. 5a and Fig. 5b. In these two respiratory challenges, the subject was switched directly from the standard mechanical ventilation to our diaphragm assist system, evaluating its ability to maintain gas exchange.
In the respiratory challenge operated with independent actuation (Fig. 5a), we see high levels of hypercarbia over time. As a result, respiratory acidosis develops, which is a direct consequence of increased PaCO2 (Extended Data Table 1a). Contrastingly, in a respiratory challenge operated with synchronized actuation in the same animal (Fig. 5b), pCO2 levels are relatively well maintained. The acidemia observed for this trial is rather of metabolic cause (called metabolic acidosis) (Extended Data Table 1b, see Supplementary Notes).
In another experiment, a respiratory trial was initiated with 2 minutes of unsupported ventilation and then switched to our diaphragm assist system, evaluating its ability to recover from a period of unsupported ventilation. During the 2 minutes of unsupported ventilation, high levels of CO2 accumulate quickly over this brief amount of time (Fig. 5c). After two minutes, the diaphragm assist system is actuated with synchronized actuation. The increasing acidification and accumulation of CO2 reverses and some recovery from the hypercarbic state is seen in the first 10 minutes, with a slight uptick in the CO2 around 15 minutes into the challenge. Complete ABGs can be found in Extended Data Table 1.
As seen by the mixed interference in Fig. 4c,h, and the ability of independent actuation to maintain blood gas balance in Fig. 5a, the alignment of the actuation with the underlying respiratory effort will critically determine the constructive versus destructive nature of the interference. In respiratory challenges that had an independent actuation scheme or a poorly synchronized actuation scheme, we found the datasets that provide a natural variation in the timing of the actuation in relationship to the underlying respiratory effort.
Because mechanical respiratory failure exists as a continuous spectrum of loss of function, we looked at the implications of synchronization in different levels of baseline respiratory effort. As seen in Fig. 3, there is variance in the underlying respiratory function between subjects. To simulate a controlled change in the underlying respiratory function within the same subject, we severed the phrenic nerve in some subjects, simulating diaphragm paralysis in combination with the respiratory depression due to the isoflurane (see Methods). Fig. 6 depicts the analysis of aligning the actuator synchronization to the underlying respiratory effort for two respiratory challenges within subject B: (1) the subject with preserved diaphragm function (Fig. 6, left) and ( 2) the subject with a severed phrenic nerve (Fig. 6, right).
To optimize for maximum inspiratory augmentation, we investigate the relationship of the timing of different waveform features to the resulting tidal volume and peak inspiratory flow of each breath. The high frequency sampling of our data acquisition system (1000 Hz) allows for millisecond temporal resolution. Custom software was written to analyze the actuation pressure, flow, and volume data.
We identify the breath bounds as determined by the local minima in the volume waveform (the locations of V0), and then finds the time distance between identified waveform features for each individual breath (further details in Methods). Waveform features analyzed include the start of an actuation waveform (P0), peak inspiratory flow (Fpk), the start of inspiration (V0), the start of expiration (Vpk), and others (Fig 6a,b, and Fig. S3).
The distances between features act as different metrics of alignment and elucidate what factors are important to consider in optimizing synchronization. There are many different features and feature distances that can be analyzed. Fig. 6c-f shows the time relationship of the start of expiration to the actuation pressure (Vpk-P0), but other metrics are shown in Fig. S3.
We examine the influence of these time metrics on tidal volume and peak inspiratory flow. We find the most important predictor variables are time metrics related to the start of expiration (Vpk). With diaphragm function preserved, there is a weak linear relationship between Vpk-P0 and the peak inspiratory flow (R 2 = 0.31, p<0.001) (Fig. 6c), and no significant relationship to the tidal volume (R 2 =0.04, p=0.001) (Fig. 6e). However, when the diaphragm function is removed by severing the phrenic nerve, a clear linear relationship emerges between Vpk-P0 and tidal volume (R 2 = 0.84, p<0.001) (Fig. 6f) and a weaker relationship with peak inspiratory flow (R 2 = 0.30, p<0.001) (Fig. 6d).
Notably, we do not find these relationships when using the timing between the start of actuation and the start of inspiration (P0 -V0) as a metric. There is no linear relationship between P0-V0 and the peak inspiratory flow or tidal volume for both the cases with and without diaphragm function (Fig. S4)
To compare the respiratory biomechanics of different modes of respiration and ventilation, pleural pressure (Ppl), abdominal pressure (Pab), and transdiaphragmatic pressure (Pdi; Pdi = Pab-Ppl) waveforms are analyzed. Transdiaphragmatic pressure is a metric of diaphragm function 6,30,31 . Pleural pressure and abdominal pressure are approximated by a sensor mounted on a balloon catheter placed in the esophagus and stomach, respectively. As these sensors approximate Ppl and Pab, the measurements are interpreted as relative measurements and not absolute measurements (see Methods for information about instrumentation and normalization). When analyzing relative pressure waveforms, the most informative metric is the maximum change in pressure per each breath.
In Fig. 7a-c, we show that across subjects (subject C was not instrumented for pressure measurements, and is therefore not shown), actuator assisted ventilation more closely matches the respiratory biomechanics of spontaneous respiration than in mechanical ventilation. Mechanical ventilation pushes air into the lungs, increasing pleural pressure with inspiration, whereas both actuator assisted ventilation and spontaneous respiration generate a negative pleural pressure to drive airflow. As the diaphragm is passive in mechanical ventilation, we see a negligible change in the abdominal pressure, whereas the caudal movement of the diaphragm in both actuator assisted ventilation and spontaneous respiration increases abdominal pressure.
In the representative waveforms from subject A (Fig. 7d-f)-the case of highest responsiveness as seen per Fig. 3c-e-the actuator assisted ventilation not only more closely resemble that of spontaneous respiration, but also augments all of the pressure waveforms. Actuator assisted ventilation generates more negative changes in pleural pressure, greater increases in abdominal pressure, and ultimately greater increases in transdiaphragmatic pressure per breath.
A graphical technique used to measure work of breathing (WOB) is the Campbell diagram, referencing pleural pressure with lung volume. Using the pressure and volume data from subject A, we generate the pressure-volume (PV) loops of a Campbell diagram (Fig. 7g). Work of breathing is calculated from this PV loop as the internal area between the inspiratory edge of the loop and the passive chest wall compliance derived from the mechanical ventilation PV data. Normal WOB is 0.35-0.7 J/L. 22,32,33 During attenuated spontaneous breathing, the subject's WOB is 0.10 J/L. During actuator assisted ventilation, the assist system shares the WOB and increases the total average WOB to 0.17 J/L, a 66% increase.
In this work, we use pneumatic soft robotic actuators to support and augment respiration, demonstrating acute augmentation of physiological metrics of respiration, and feasibility as a proof-ofconcept device. A set of two McKibben-style PAMs surgically implanted superior to the diaphragm are capable of providing mechanical support to the diaphragm in a large animal model of respiratory insufficiency. We thoroughly characterized the in vitro mechanical properties of the device and investigated its interactions with the respiratory system and the subject, using multimodal metrics to evaluate respiratory function (e.g. tidal volume, inspiratory flow), biomechanics (cavity pressures, WOB), motion (ultrasonography and fluoroscopy), and gas exchange (ABGs).
The diaphragm assist system generated substantial augmentation in respiratory function-measured via peak inspiratory flow (a direct metric of inspiratory function), and tidal volume and minute ventilation (metrics of ventilation)-in our most responsive subject. Subject A had the highest change in peak inspiratory pressure, tidal volume, and minute ventilation; the corresponding large augmentation in peak inspiratory pressure indicates that the volume and minute ventilation augmentation are specifically due to the soft robotic actuators augmenting the diaphragm's inspiratory function. Responsiveness to the system varied across subjects.
Variance in responsiveness is likely dependent on a combination of many factors. One factor is the level of preserved respiratory baseline. The weak response in the subject with a relatively high preserved weight-normalized minute ventilation (subject E) suggests that the assist system may have weak augmentation or even a disruptive effect in cases of well-preserved diaphragm function. Other potential factors include precise actuator placement, actuator fit, and anatomical variations.
We showed that synchronization with the native respiratory effort is a critical design element in our system. Synchronous actuation is key to consistent, low-variance respiratory waveforms and tidal volumes. Like standard mechanical ventilation, off-cycle actuation of the actuators can lead to a destructive interference with the underlying respiratory effort, resulting in a poor augmentation and poor blood acid-base balance. In evaluating the effect of synchronization on the system's ability to maintain appropriate gas exchange, we demonstrated that despite generating a similar range of tidal volumes, independent actuation led to an inability to maintain appropriate pCO2 levels and resulted in respiratory acidosis. Contrastingly, in two trials of well-synchronized actuation, we observed some capacity of the device to maintain and recover baseline pCO2 levels.
The control system used in this study was a simple but effective first-generation system with many directions for improvement. The synchronization triggered from airway flow-which is also the metric used by gold standard clinical ventilatory support options for triggering-but flow is also the most downstream signal in neuro-ventilatory coupling. The downstream nature of the signal is a potential source of delays and asynchrony 34 . In order to achieve consistent assistance from breath to breath, the synchronization must be optimized for the alignment that maximizes constructive interference. The system relied on a manually titrated threshold set for the flow sensor data. It is designed to be triggered at the start of an inspiratory flow effort, which is related to V0. However, the manual nature of the system meant that if the threshold was set too low, noise in the flow signal could cause preemptive or false triggering (as evidenced by the negative values for P0-V0). Our alignment analysis reveals two important considerations for improvements towards this goal. The first consideration is that the influence of alignment changes with the degree of preserved respiratory function, as seen with the difference in results between the intact and the severed phrenic nerve. When the phrenic nerve is severed, all diaphragm motion is governed by the actuators, and misaligned actuation with the remaining native respiratory effort-expansion of the ribcage-results in more consequential destructive interference. Whereas when the phrenic nerve is intact, the net diaphragm motion results from a combination of native diaphragm function and the effect of the actuators, because the actuators only operate along 2 discrete lines on the diaphragm. The contraction of the rest of the native diaphragm motion is still synchronized with the ribcage motion, so the effects of misalignment are less apparent. This implies that optimal alignment parameters may be different for different disease states and the control system will need to be dynamic and adaptive to changes in respiratory function, even within the same patient. The second consideration is that the actuation curve's relationship to the beginning of expiration (Vpk) is more influential than the relationship to the beginning of inspiration (V0). This implies that an updated system should trigger from a signal related to expiration as opposed to the beginning of inspiration. Some neuromuscular signals, like the electrical activity of the diaphragm (Edi), contain detailed information about both inspiration and expiration times 35,36 . Edi amplitude is also proportional to the neural drive, as well as the degree of contraction of the diaphragmatic muscle, therefore opening up the possibility of adaptive control. Triggering from Edi measured at the esophageal level via a feeding tube 37 may be warranted to improve mechanical ventilation. This method, known as neurally adjusted ventilatory assist, is available in the clinical setting with mechanical ventilation and may improve respiratory weaning of patients that are challenging to wean 36 . The same principle could be applied to our diaphragm assist system; using a more upstream signal with greater information on the native respiratory effort would allow for a more robust control system.
Overall, we show that the strategy to augment the native function of the diaphragm with soft robotics acts as a form of negative pressure ventilation by driving ventilation through the generation of a negative pressure in the thoracic cavity. Our diaphragm assist system is biomechanically similar to that of spontaneous breathing, sharing a substantial portion of the work of breathing in our best responding subject. By functioning as an assist device-as opposed to completely overtaking breathing-our system has the potential to be compatible with voluntary use of the diaphragm. Maneuvers such as voluntary deep breaths or drinking through a straw-abilities related to patient autonomy and quality of life-can be preserved with this implantable ventilator strategy. Additionally, in contrast to current modes of mechanical ventilation, recapitulation of native biomechanics, as shown with this system, can avoid the deleterious effects that arise secondary to the use of positive pressure ventilation, such as barotrauma 38,39 or hemodynamic changes in patients with concurrent cardiac pathologies 40,41 .
In this study, we demonstrate the foundational work towards a soft robotic implantable ventilator. Translationally, there are many hurdles to overcome between the proof-of-concept state presented here and the ultimately envisioned system, and we discuss them in the subsequent text.
Given that we saw variable responsiveness to the device across subjects, additional studies are needed to understand what factors in system design and implantation can replicate high responsiveness. Our system could generate the low end of acceptable minute ventilations but relied on high respiratory rates to do so. Given the presence of dead space, low tidal volumes result in less alveolar ventilation than if the same minute ventilation achieved with higher tidal volumes and a lower respiratory rate. A core goal of the next generation system is to further improve the tidal volume augmentation, which will need to be achieved through both actuator design and control system development.
Here, we used the classic McKibben actuator; a more application-specific or customized actuator type may allow for further increases in tidal volumes in future work. Other factors in actuator design, such as the number, layout, and positioning of actuators, will also be critical. We demonstrated tunability of assist by controlling pressurization, but an updated design will require finer characterization. Synchronization is critical to device performance, and thus future work lies in building a next generation control system; this includes creating a system that is cognizant of the beginning of expiration as opposed to inspiration, an automated control system that removes the error of manual titration, and further investigation of dynamic actuation curves. An ideal next generation control system aims to trigger from a more upstream neural signal-such as the electrical activity of the diaphragm-to provide an earlier signal that enables an advanced control system to optimize synchronization, removing delays and asynchrony. Neural triggering via implanted electrodes would also untether the current system from the flow instrumentation, freeing the patient from interventions at the mouth or trachea. To fully realize untethering from bulky machines-like standard mechanical ventilators-the external components that control and power the system require miniaturization. Future work will aim to eventually miniaturize the system to the scale of a small backpack-one that could be worn by the patient or attached to a belt or an electric wheelchair. The process of miniaturization and portability has proved to be possible in similar complex devices, such as ventricular assist devices (e.g. Thoratec HeartMate III) or total artificial hearts (e.g. Syncardia TAH, Carmat Aeson) [42][43][44][45] .
Envisioning a translation to the clinical field, the following considerations might help to optimize the management and pave the way to human application. The diseases leading to chronic diaphragmatic dysfunction are numerous and feature very different pathophysiologies. Therefore, a thorough understanding of the underlying pathology as well as its specificity are critically needed to help optimize the management and anticipate complications 46 . Moreover, patient selection and indication will need to be clearly defined, in order to select the patients who will benefit from this therapy the most. Here, we present a generalized mechanical strategy for diaphragm support, but the parameters of actuator design or actuation control will need to be optimized and specialized per the needs of a given pathology as well as individual patient anatomy.
Owing to the complexity of the procedure, a multidisciplinary team highly trained in advanced thoracic surgery is required to build expertise and develop this technology, ideally in a high-volume center 47 . Technological improvement is required to provide the least invasive approach of implantation. In this regard, a thoracoscopic route might be beneficial and will be the subject of future work. Given the invasive nature of implantable devices, the diaphragm assist platform is targeted towards patients with chronic-to-permanent ventilator dependence. We recognize that surgery in patients suffering severe diaphragm dysfunction causing respiratory failure can carry a high morbidity and mortality. Perioperative complications can be numerous; one of the most feared is the worsening of the pulmonary status, which may itself precipitate the need for long-term ventilation 48 . Nevertheless, it has been well demonstrated that complex thoracic surgery is feasible even in very frail patients. Lung transplantation for terminal respiratory disease 49 is one of the most striking examples. Thus, surgery could still be considered in a suitable target population that would ultimately benefit from this mechanical augmentation of diaphragm function, such as a range of neuromuscular disorders. The concept of diaphragm assist is in itself a means of preventing further complications from chronic respiratory failure and preserving key aspects of quality of life-like speech and mobility.
Due to the focus on feasibility, we acknowledge that there are limitations in these acute studies from the lens of regulatory approval and clinical translation. We do not study device biocompatibility or long-term device operation. The device was constructed from types of polymers that are already used in established medical devices [50][51][52][53] , such as poly(ethylene terephthalate) (PET) and polyurethanes (See Supplemental Information). Because our device focuses on mechanical interaction, as opposed to biochemical interactions with the body, the materials used in the device can easily be substituted with regulatory-approved materials in future iterations. With improved performance and stability, future long-term studies will need to investigate the long-term effects of the system including tissue remodeling and the ability to provide full-time respiratory support.
Although this technology requires further advancements in the net tidal volumes it can generate before it can fully match the ventilation capacity of a current mechanical ventilator, it is the first study to report the ability to rescue ventilation with an implantable ventilator. We envision further translational potential of this technology when combined with the development of smaller and more portable pneumatic energy sources 54,55 as the field of soft robotics advances. With the integration of a portable pump and control system in the future, this technology could provide an additional level of patient autonomy via increased mobility. Motivated by the encouraging results of this study, we believe this technology, with optimized design, has the potential to provide a radically different ventilation technology that preserves key metrics of quality of life for people with end-stage mechanical respiratory failure.
There were two main objectives of our study. First, we sought to demonstrate the proof-of-concept capability to augment ventilation via implanted soft robotic actuators in an animal model of respiratory muscle weakness. To evaluate ventilation metrics, we measured spirometric flow and volume. Second, we aimed to demonstrate that this soft robotic strategy replicates more native respiratory biomechanics than standard mechanical ventilation. To evaluate the respiratory biomechanics, we evaluated the respiratory pressure data along with the spirometry data.
In order to evaluate the system performance under varying conditions within a single animal, a series of respiratory challenges were performed. Prior to the first and between each subsequent respiratory challenge, volume control mechanical ventilation operated through the facility's Drager Tiro ventilator (Drägerwerk AG, Lübeck, Germany) was used to maintain the animal's ventilation needs and recover from respiratory challenges if needed. Arterial blood gases were taken to validate normal baseline respiratory status prior to each challenge. Each respiratory challenge was initiated by switching the ventilator to a manual mode of ventilation. Data for a mix of unsupported ventilation and actuatorsupported ventilation was collected. Vital signs and respiratory status were monitored. For experiments with uninterrupted trials, ABGs were collected at 2 or 5-minute intervals during the challenge.
Fabrication and characterization of pneumatic artificial muscles. The actuators were a modified version of the PAM actuators previously described in Roche, E.T. et al 7 , and Payne, C., et al 9 . Specifically, McKibben pneumatic artificial muscles were fabricated according to the protocol detailed in the Supplementary Methods. Actuator dimensions were selected to fit the anatomical needs of the 30-40 kg swine. They consist of a thermoplastic elastomer bladder (Stretchlon 200, FibreGlast Developments Corp.), a thermoplastic polyurethane tubing (1/8" Tubing, 5648K226, McMaster, Inc) and an expandable braided mesh (PTO0.25BK, TechFlex, Inc.). Prior to use in vivo, actuators were fatigue tested to a pressurization of 20 psi for >1000 cycles on the benchtop. Mechanical characterization was performed on an Instron 5499 Universal Testing System (Instron Corp, Norwood, MA, USA).
Actuator characterization was conducted both in vitro and in vivo. For the in vitro characterization, actuator performance was measured via Instron testing. Classic tensile testing was conducted to measure the contractile force. A modified flexural bend setup (Fig. S5) was used to measure the perpendicular force applied to the diaphragm via arclength shortening. For the in vivo characterization, performance of the diaphragm assist system was evaluated through the diaphragm displacement (via ultrasonography) and the functional metrics (tidal volume, Campbell diagram) (Extended Data Fig. 1 and 2). Different pressurization shapes and levels were input into the actuator (Extended Data Fig. 1 and 2) and the resulting behavior was measured. Further details can be found in the Supplementary Information.
All studies were conducted according to protocol #19-05-3907 approved by the Boston Children's Hospital (BCH) Institutional Animal Care and Use Committee (IACUC) policy.
Procedures were carried out at Boston Children's Hospital in accordance with BCH IACUC under protocol #19-05-3907 and MIT IACUC under protocol #0118-006-21. Protocol reviews were conducted in accordance with the standards outlined in the National Research Council's Guide for the Care and Use of Laboratory Animals and BCH's Animal Welfare Assurance. Female Yorkshire 30-40kg swine were sourced from Parson's Farm (Hadley, MA, USA). We used a total of twelve swine during the development and testing of our system, and we present data from nine swine in the manuscript. Different subsets of subjects were used for the experimental investigations reported; not all subjects were used in every experimental investigation. Animals were acclimated and cared for according to standard facility protocols. Each experiment was conducted under 2-3% isoflurane anesthesia, titrated to each animal to maintain a stable anesthetic plane. Anesthesia and mechanical ventilation were controlled through the facility's Drager Tiro ventilator (Drägerwerk AG, Lübeck, Germany). Vital signs were monitored via a SurgiVet monitor (Smiths Medical, Inc., Minneapolis, MN, USA). After completing the study and acquiring the data, animals were euthanized using Fatal-Plus Solution (Vortech Pharmaceuticals, Dearborn, Michigan) at a dose of 110 mg/kg/body weight.
After induction of anesthesia, the animal was intubated and placed on mechanical ventilation. A transesophageal EKG catheter was placed to monitor the heart rate. A carotid arterial sheath and jugular venous line were placed using cut-down technique for animal systemic and central venous pressures monitoring respectively. Two balloons were placed, one in the esophagus and one in the stomach for pressure monitoring. A Foley catheter was placed for urine output monitoring. Following that, the chest cavity was accessed through midline sternotomy. Next, we opened both pleural cavities and placed one soft actuator along the diaphragm curvature on each cavity. The anterior portion is attached to the sternum and the posterior attachment is made to the lowest posterior rib in the most medial position that can be achieved without disrupting the region of the major arteries and veins, esophagus, and spine. To do this, we passed each actuator posteriorly at lowest intercostal space to outside the chest cavity and fixed it to the skin using sutures. Then, we fixed the other end to the sternum using sutures and we passed the actuation lines through separate opening through the skin. Next, we approximated the sternum using sternal wires and close the subcutaneous layers and the skin in layers using sutures. After the sternotomy was closed, the negative pressure in the thoracic cavity was restored via chest tube, and the respiratory challenges were conducted with a closed chest.
In order to simulate varying levels of respiratory functions, two animal models of respiratory muscle weakness were used. The first method relied on the respiratory depressive effects of isoflurane. Isoflurane levels were held between 2-3% and titrated to a stable plane of anesthesia while still maintaining a depressed but non-zero level of spontaneous respiration during respiratory challenges. The second method modeled diaphragm paralysis via mechanically severing both the left and right phrenic nerves. This model is still conducted under the setting of the isoflurane, and thus combines the effects of the isoflurane and severed phrenic nerve and represents a more severe model of respiratory model weakness.
The biomedical sensors and instrumentation data were input into a PowerLab 35 series (PL3516, ADInstruments, Dunedin, New Zealand) high performance data acquisition system with a 1000Hz sampling frequency for all channels. During the experiments, data was monitored live via LabChart software (ADInstruments, Dunedin, New Zealand). After the experiements, data was exported into and processed in MATLAB (MathWorks, Portola Valley, CA, USA).
An analog spirometer (Gas Flow Sensor, ES Systems, Filothei-Psychiko, Greece) was placed in line between the ventilator Y-tubing and the endotracheal tube. Analog data was input into the PowerLab. The data was converted from mass flow to volumetric flow according to manufacturer specifications Respiratory pressure instrumentation and measurement. Pleural pressure and abdominal pressure were measured via an esophageal balloon catheter (Cooper Surgical, Trumbull, CT, USA) placed in the esophagus and stomach respectively, each connected to a pressure transducer (PRESS-S-000, PendoTech, Princeton, NJ, USA).
Respiratory pressure data was normalized in the MATLAB post-processing. For a given segment of interest, the average of the pressure reading at the breath bounds was set to zero, to allow the analysis to show the change in pressure over the course of one breath.
Ultrasonography (US), a non-invasive, non-ionizing imaging method, was used to investigate the interactions of the device with the diaphragm. US can be used to assess diaphragm displacement and dysfunction 25 . More precisely, it allows a direct two-dimensional visualization of the diaphragm, permitting a quantification of its motion and function, and serves as an ideal tool to assess the interaction of the device with the diaphragm. A Philips iE33 (Philips Healthcare, Andover, MA, USA) echography machine was used with the X7-2 transducer (Philips Healthcare, Andover, MA, USA). A 2dimensional image (so-called B-mode, Brightness) of the diaphragm and the device was obtained by placing the probe in the right subcostal space, pointing in the cranial direction. To quantify the motion of the device and the diaphragm, M-mode was used.
Our group has built a custom electropneumatic control system utilizing electropneumatic pressure regulators and valves (SMC Pneumatics, SMC Corp, Tokyo, Japan) controlled by a custom software described in Horvath, Hu et al 56 . The software is designed to allow custom pressure waveforms to be input. The control system can generate a desired waveform via an analog input to the electopneumatic regulators. The nominal peak pressure for all waveforms was 20 psi. The regulators also output an analog signal of the actual pressure waveform; this data is input into the PowerLab system.
Independent and synchronized actuation. The custom control system is capable of generating a manual timing set to a frequency of actuation that is initiated by user input. This set timing initiates the custom pressure waveform programmed into the system. This timing is independent of the subject's native breathing.
In order to implement synchronization in our system, the Fast Response Output add-on for LabChart (ADInstruments, Dunedin, New Zealand). Analog spirometry flow data was used as the input channel. Voltage and hysteresis settings were manually titrated between a voltage range equivalent to 0.01 L/s to 0.07 L/s and a hysteresis range between 2-5% during every respiratory trial to achieve qualitatively good synchronization, as visually recognized by the homogeneity of the real-time flow and volume waveforms. The digital output channel on the PowerLab was used to send a trigger pulse to a digital input channel in the microcontroller of the custom control system described above.
Statistical tests were conducted as described in the respective figure captions for Fig. 3, 4, 7, and Extended Data Fig. 1 and 2. For Fig. 3c,d and Fig. 7a-c, two-sided Wilcoxon rank-sum analyses were conducted in MATLAB (MathWorks, Portola Valley, CA, USA) via the "ranksum" function. Fig. 4e,f depicts two sets of statistical tests. A two-sided Welch's t-test without assuming equal variances was conducted in order to compare the means of the populations via the "ttest2" function in MATLAB with an "unequal" variance type specification. Additionally, a 2-sample F-test for equal variances was conducted to compare and confirm unequal variances via the "vartest2" function in MATLAB. For the Extended Data Fig. 1 and 2, two-sided t-tests were conducted via the "ttest2" function in MATLAB. 1. Full arterial blood gas results. Arterial blood gas results for the respiratory challenges reported in Fig. 5. Here, Table 1 a,b,c corresponds with Fig. 5a,b,c respectively. Time [s] 0 370 618 pH 7.449 7.307 7.294 pCO2 [mmHg] 37.8 89.2 68 HCO3 [mmol/l] 26.4 45 33.3 pO2 [mmHg] 313.4 228.2 189 sO2 [%] 99 100 99 Time [s] 0 262 626 pH 7.408 7.332 7.285 pCO2 [mmHg] 47.4 44.9 47.2 HCO3 [mmol/l] 30.2 24 22.6 pO2 [mmHg] 246 217.9 158.2 sO2 [%] 99 100 100 Time [s] 0 124 433 732 1010 pH 7.471 7.337 7.359 7.419 7.389 pCO2 [mmHg] 35.6 75.1 63.2 33.6 54.7 HCO3 [mmol/l] 26.2 40.6 36 21.9 33.4 pO2 [mmHg] 256.6 233.1 292.4 232.7 272.2 sO2 [%] 99 99 100 100 99
Actuator performance was characterized both in vitro and in vivo as described in the Methods in the main manuscript and the Supplemental Methods presented below.
Different input shapes explore the effect of rate of pressurization (Extended Data Fig. 1a-c). The fidelity to these idealized waveforms is limited by the control resolution of the electropneumatic regulators, and ultimately result in the output pressurization curves of Extended Data Fig. 1d-f. These actuation pressure curves ultimately govern the mechanical performance of the actuators. The actuators are characterized in vitro via tensile and flexural testing, as described in the Supplemental methods. The tensile force (Extended Data Fig. 1g-i) represents the contractile force applied to the points of attachment on the ribs, and the flexural force (Extended Data Fig. 1j-l) represents the force perpendicular to the actuator towards the diaphragm.
Different actuation pressure waveforms result in different displacements, (seen in the M-mode ultrasound in Extended Data Fig. 1m-o and quantified in Extended Data Fig. 1p), tidal volumes (Extended Data Fig. 1q) and different respiratory mechanics (Extended Data Fig. 1r). Notably, the square wave pressurization is distinct from the behavior of the curved wave and triangle wave, especially with regards to the average diaphragm displacement and the Campbell diagram. We note that the square wave achieves similar tidal volumes to the other waveforms while drawing more negative pleural pressures. The slope of the Campbell diagram, taken at the two points in the loop where 𝑑𝑉 𝑑𝑃 = 0, can be viewed as a representation of compliance of the system. A negative pleural pressure drives flow via the gradient from atmospheric pressure at the airway opening to the negative alveolar pressure, so we evaluate the absolute value of compliance. The slope generated by the square wave (12.1 mL/cmH20) is considerably lower (i.e., the system is stiffer) than that of the curved (15.9 mL/cmH20) and triangle wave (15.1 mL/cmH20), which both have slopes that more closely resemble those that of spontaneous respiration (25.6 mL/cmH20). These values are overall relatively stiff and are likely due to the low lung volumes for this subject. Qualitatively, we observe that the square wave results in "sharper" breaths that pull on the chest wall more aggressively compared to the gentler inflation of the curved and triangle wave, matching the much higher tensile forces generated by the square wave from in vitro testing.
The curved and triangular pressurization input are similar in their pressure-volume (PV) loops; however, the curved input achieves higher tidal volumes with marginally smaller levels of diaphragm displacement (Extended Data Fig. 1p-r), which could be attributed to the shorter time in which the actuator operates at a high pressure filled state which may not provide enough time for lung filling. Overall, the curved waveform used in the majority of the study (Fig. 234567in the main text) represents a pressurization scheme that aims to combine the benefits of the square wave and triangular wave, generating the best tidal volumes and biomimetic PV loops.
PAM performance can also be tuned via depth of pressurization by scaling the input curved waveform shown in Extended Data Fig. 2a to different peak pressures (5, 10, 15, and 20 psi), the resulting actuator pressure waveforms are shown in Extended Data Fig. 2a-d. The relationship between pressurization and forces generated is linear (Extended Data Fig. 2e,f) which corroborates previous McKibben characterization work 24 . We characterize the response of one subject to the varying degrees of pressurization. The degree of pressurization has a positive, but nonlinear effect on the amount of diaphragm displacement generated (Extended Data Fig. 2g-k). In this subject, we demonstrate tunability of the degree of augmentation via changes in pressurization, with the greatest range of responsiveness being between 0 and 10 psi. Additional increases taper off between 10 and 20 psi, which matches the understanding of how McKibben actuators operate, as they first expand and fill to their maximum volume, achieving maximum contraction, and beyond that they increase force generation 24 . In terms of respiratory mechanics, the degree of pressurization does not have a large effect on the change in pleural pressure (Extended Data Fig. 2m) unlike the different waveform shapes in Extended Data Fig. 2p.
Notably, interanimal variability is undeniably a factor contributing to overall performance, as evident in the varied responsiveness to the device seen in Fig. 3. Even in a case of low augmentation, we saw a nonlinear but tunable response to different levels of pressurization. The absolute degrees of augmentation shown in Extended Data Fig. 2k-m will obviously not hold across different animals, but we expect that the relative effect of tuning pressure should.
Extended Data Table 1 depicts ABGs taken during the respiratory challenges depicted in Fig. 5 with (a) independent actuation, (b) synchronized actuation, and (c) synchronized actuation after 2 minutes of unsupported ventilation. During (a), we observe a progressive blood acidification, a consequence of increasing PaCO2 (respiratory acidosis). In parallel, HCO3 increases, and is an adaptative mechanism of the organism to correct pH value. However, this compensatory mechanism only partially helps controlling pH. During synchronized actuation (b), PaCO2 remains constant around the upper level of the normal range. We see a progressive acidification of the blood, whose main primary cause is a decrease in HCO3 (metabolic acidosis). For respiratory challenge (c), ABGs shows a clear respiratory acidosis (elevated PaCO2) with partial compensatory mechanism (increase in HCO3) during the unsupported ventilation phase. While actuating, PaCO2 decreases, as well as HCO3, resulting in a controlled pH. At the end of the trial, P a CO 2 slightly increases with a corresponding HCO 3 compensation. PaO2 and O2 saturation (SaO2) remained in the normal range across all experiments.
As discussed in the main manuscript in the section entitled "Factors in optimizing synchronization", there are many different alignment factors that can be analyzed (Fig. S3). One additional metric of interest is the relationship between the start of actuation (P0) and the start of inspiration (V0). Because the triggering system presented here relies on the beginning of inspiration for triggering, this is a metric that we hypothesized would be influential. We find no linear relationship between (P0-V0) and the tidal volume or the peak inspiratory flow (Fig. S4). Although we see a clustering effect in Fig. S4a,c, within the clusters, there are no linear relationships. The clusters themselves can be understood as two categories of actuation. Actuation that occurred early (likely falsely triggered) with a negative P0-V0 value and actuation that occurred in an appropriately synchronized fashion. With this distribution of behavior within one respiratory challenge, the data actually resembles the data from comparing