The dawn of physiological closed-loop ventilation—a review

An automated closed-loop system (also known as feedback control) can be implemented to keep a patient at a specified target and respond to disturbances without the clinician’s presence being necessary. Hereby, a controller takes over the task of adapting ventilator settings. The new control loop is depicted in Fig. 2 and shows the subtraction of a measured value from the target value and the error being fed into the controller. The controller then automatically decides on the correct ventilator settings to minimize the error.

Importantly, the clinician’s focus changes to choosing personalized targets, regulating variables supplementary to the ventilation, such as hemodynamics and fluids, and monitoring the system. It should be noted that the current PCLC systems limit their measurement data to classical ventilation variables, as will be shown later. Derived evaluations, such as pF ratio, stress and strain, or heart-lung interactions, have not yet been used.

The goals of an automated closed-loop system can be divided into setpoint tracking and disturbance rejection. For most cases, setpoint control is only relevant when the controller is switched on, as it describes the dynamic response of the system until the target is reached. Setpoint changes are rare in the clinical environment. Exemplarily, the SpO₂ target may be at 88–95% and seldom changes. The disturbance rejection, however, is the true hero during PCLC, as it is concerned with the response of the system to disturbances and bringing the patient back to the target range. Disturbances can take various forms, both internal (worsening illness, lung stiffness, increased CO₂ production) and external (disconnection, room temperature). The two goals are shown for an illustrative example in Fig. 3, with a setpoint change at t₁ and a disturbance (extreme increase in CO₂) at t₂.

The evolution of automated physiological control can be categorized by two drivers: the control target and the controller design. The control target is dictated on the one hand by the measurements and sensors available, some of which have been described above. On the other hand, changing clinical evidence and guidelines present new control targets. The second driver, controller design, remains an engineering problem and is beyond the scope of this paper. A brief overview is given here only to facilitate understanding. A natural evolution of the complexity of the controllers comes with the use of computers and their increasing computing power, and the detailed modeling of the respiratory system, which is used for controller synthesis and testing. Simple proportional, integral, and derivative (PID) controllers were used initially [15, 16]. This summation of PID action is the most common controller used in almost all industries nowadays. The tuning (setting up of this controller) is static, meaning the controller needs to be tuned to every patient. To overcome this, adaptive controllers [17‐19] can be used, which change their controller parameters to adapt to the patient or scenario. A mathematical model of the system is required for this, but recent advances in modeling respiratory systems have allowed their increased use. The optimal controller attempts to solve an optimization problem, such as minimizing a cost function, in calculating the actuation variable [20]. Model predictive control (MPC) is another example of an advanced controller which uses a dynamic model of the system [21, 22]. It optimizes the current ventilator settings for the current state, while anticipating future events. Finally, there has also been increasing use of fuzzy control [23, 24]. A fuzzy controller is a rule-based control law, in which the linguistic rules (e.g., if-then-else rules) are fuzzified (blurred) in order to make the control more elastic. These systems are popular in the clinical environment, as the language-based expert knowledge of clinicians is transferred into the design of the controller (which is not the case in PID control, for example). More robust controllers may be introduced in the future. These are designed with uncertainty in mind and can control the performance of the system even for the worst case.

The first application of PCLC in mechanical ventilation was presented by Saxton in 1953, with a publication appearing in 1957 [15], where his team applied feedback control to the iron lung to regulate the etCO₂. In the same decade, Frumin developed an automated anesthesia system, which incorporated an etCO₂ feedback control system [16, 25]. Both systems used the ventilation pressure as the actuating variable and were able to keep end-tidal CO₂ at the set target.

In 1971, Mitamura and colleagues controlled the mixed expired CO₂ using both tidal volume (V_T) and breathing frequency (f) [20]. Their system was able to keep PCO₂ at the target, even with extracorporeal CO₂ loading. Other groups focused on setting either V_T or f automatically, with the clinician setting the other variable manually [26, 27]. Digital control using computers for CO₂ control started with Coles et al. in 1973 [26], and the number of publications about feedback control in ventilation increased. Coles et al. showed that the PCLC system maintained the etCO₂ at the target better than manual control [26].

The availability of intravascular sensors for pH or PaCO₂ measurement introduced new closed-loop control systems. Schulz et al. used such a sensor and feedback control to respond to setpoint changes in PaCO₂ [28]. The response dynamics of their system was, however, limited by the slow response time of the sensor. A similar sensor problem caused the system of Coon et al. [29] to oscillate. The continuous intravascular sensors, however, failed to remain commercially available and their use has ceased. Without any clinical alternatives, the control of CO₂ was again based on the measurement of etCO₂. However, the increasing difference between PaCO₂ and etCO₂ in the pathological lung requires compensation [20, 28, 30]. Approaches based on first principles to link etCO₂ and PaCO₂ were presented by Ohlson et al. [30], but the authors could not compensate correctly when a large variation of cardiac output appeared.

An extensive list of CO₂ feedback control (often referred to as ventilation control) systems over the past 50 years is shown in Table 1. As can be seen, all systems presented so far have been limited to proof-of-concept studies.

The focus was mostly on the control of CO₂ until the early 1970s, but in 1975, Mitamura [36] developed a “dual control system” for both CO₂ and O₂. Here, the SpO₂ was measured using an ear oximeter and an on-off controller was used for changing the inhaled oxygen content. The controller was able to rectify the hypoxia.

Controlling only the oxygenation was performed in preterm infants by Beddis et al. [37] in 1979; this was made possible by using an indwelling umbilical arterial oxygen electrode sensor. To evaluate their system, they compared the time spent at the oxygenation target using the controller to a clinician-in-the-loop system. This metric was also used by others [38‐44] and the automated system was as good or better than the manual procedure in all cases.

Yu et al. later used an oximeter on the tongue to control the oxygenation and implemented an adaptive controller in dogs [19]. This system was able to rectify hypoxia and compensate for disturbances, such as positive end-expiratory pressure (PEEP) changes and one lung ventilation. In 1991, East et al. [45] used both FiO₂ and PEEP to control PaO₂. Whether to change FiO₂ or PEEP was based on previous clinical protocols and the system kept the patients at the oxygenation target for up to 6 h [45].

Further literature on oxygenation control is summarized in Table 2. The control of oxygenation using the fraction of inspired oxygen remains an active field of research, especially in neonates. A recent review of oxygenation control was published by Claure and Bancalari in 2013 [46]. The systems presented by Claure et al. [41], Urschitz et al. [42], and Gajdos et al. [44] have been developed further and are now commercially available as AVEA-CLiO2 (CareFusion, Yorba Linda, CA, USA), CLAC (Löwenstein Medical GmbH & Co. KG, Bad Ems, Germany), and SPOC (Fritz Stephan GmbH, Gackenbach, Germany), respectively.

Importantly, the control of oxygenation has mostly been limited to using only the FiO₂ so far. This closely reflects the clinical difficulty in correctly choosing the PEEP—not least because oxygenation alone is not a reliable measurement of a good PEEP.

A large disadvantage of most of the closed-loop ventilation strategies presented is that they focus only on gas exchange and do not consider lung mechanics and quantification of harm. In fact, achieving proper CO₂ control may require excessive V_T or peak inspiration pressure, which can cause ventilator-induced lung injury (VILI). With clinical ventilation strategies becoming focused on being protective and preventing VILI, this requirement also needed to be incorporated into closed-loop control. Hence, control considering lung mechanics is presented next.

Mitamura et al. considered minimizing ventilatory work as a further goal of their controller as early as 1971 [20]. The idea is closely related to that of Otis et al. [50] from 1950, which suggests that there exists an optimal combination of respiratory rate and tidal volume for minimal work of breathing. This approach was also used by Tehrani [51] and Laubscher [52] in 1991 and 1994, respectively. Laubscher et al. showed that their controller was able to adapt to personalized respiratory mechanics in a study on six patients. Laubscher and colleagues advanced their adaptive lung ventilation (ALV) controller (1994) to a newer version called adaptive support ventilation (ASV). Arnal et al. [53] tested the ASV controller on 243 patients with different respiratory lung conditions and showed the ability of the controller to choose V_T– f combinations related to actual personalized lung mechanics.

Rudowski et al. [54] addressed the concerns of VILI directly with the peak respiratory power index, as an index of lung trauma, in 1991. Their controller adjusts ventilator settings to reduce the respiratory power index, while ensuring adequate gas exchange. A study with six patients showed promising results.

Many modern systems do not directly control the ventilator using lung mechanics, but rather apply hard limits to ensure that tidal volume and pressure stay within certain, evidence-based, limits.

It is also important to note here that the majority of literature presented so far considers only controlled (mandatory) ventilation, i.e., the patient is passive. However, in many cases, the patient is allowed or expected to breathe spontaneously, meaning patient demand becomes important. Early work was performed by Younes et al. with proportional assist ventilation (PAV) in 1992 [55]. This positive feedback control system amplified patient effort, according to the respiratory mechanics and level of assistance set by the operator. This ensures synchrony, while automatically adapting to patient load. For the initial version of PAV, the clinician required knowledge of the respiratory mechanics of the patient to set an appropriate controller gain, but the newer version, called PAV+, estimates the individual respiratory mechanics automatically [56, 57].

In 1996, Iotti et al. proposed P_0.1 closed-loop control ventilation, whereby the drop in airway occlusion pressure during the first 0.1 s of inspiration is used to estimate patient work [58]. Two independent controllers, one for P_0.1 and the other for alveolar volume, are fed into a merged control algorithm which changes the level of pressure support. The authors showed that inspiration activity of the patient can be stabilized at a desired level using P_0.1, thus allowing for the unloading of the inspiratory muscles.

A direct coupling to the physiological neural output of the respiratory system would be helpful for optimal support during spontaneous breathing. An attempt to couple a respirator to phrenic nerve activity was performed in 1970 on animals [59], but this was not feasible in humans. Instead, Sinderby et al. [60] used the diaphragmatic electrical activity (EAdi) for neuro-ventilatory coupling to adjust the level of ventilatory support. This system requires the placement of an esophageal catheter. This system is commercially available as neurally adjusted ventilatory assist (NAVA) (Maquet Critical Care AB, Solna, Sweden). Improved patient-ventilator synchrony for this system was shown by [61], but the authors noted that the clinical impact thereof still needed to be determined.

There have been various approaches to computerize clinical protocols which medical staff use to adapt mechanical ventilator settings. These computerized decision support systems do not make active changes to the ventilator but rather propose a change; the clinician has to be present to make the change and, as such, they represent classical clinician-in-the-loop systems. A literature review of these systems is given in [62].

Pomprapa et al. automated the ARDSNet protocol (autoARDSNet) and tested the system in a pilot study with seven pigs [63]. The system kept the animals at the oxygenation and pH targets for a duration of 4 h, even compensating for a disconnection between ventilator and patient.