Closed-loop non-invasive ventilation in critical care: current evidence and clinical perspectives

Introduction

Non-invasive ventilation (NIV) has become an essential component of respiratory support in acute, critical care, and perioperative settings. Guidelines support NIV for acute hypercapnic respiratory failure, cardiogenic pulmonary edema, and selected post-extubation or postoperative scenarios, whereas its use in de novo hypoxemic respiratory failure requires careful patient selection and close monitoring (1, 2). The Research Topic “Critical Care Focus in Non-Invasive Ventilatory Approaches: a Loop Approaches” emphasizes technological advances, monitoring, hypercapnic respiratory failure, critical care anesthesiology, and continuity of care across phases of treatment. These priorities make closed-loop and adaptive NIV an attractive, but still underdeveloped, field.

Table 1. Comparative analysis of current and emerging closed-loop strategies relevant to non-invasive ventilation in critical care.

The clinical promise of NIV is inseparable from its technical fragility. Unlike invasive ventilation, NIV operates through a non-hermetic interface, with intentional and unintentional leaks that can distort flow estimation, trigger performance, cycling-off, delivered tidal volume, and displayed monitoring variables (3–5). NIV also preserves spontaneous breathing. This can maintain diaphragm activity and avoid endotracheal intubation, but may expose the lungs and respiratory muscles to injurious effort if support is insufficient or if treatment failure is recognized too late (6–8). In this context, the ideal NIV system would not merely deliver a prescribed pressure; it would continuously sense leaks, respiratory timing, mechanics, gas exchange, inspiratory effort, and clinical trajectory, then adjust support while preserving clinician oversight.

This mini-review focuses on closed-loop and adaptive strategies for NIV in critical care. Because fully autonomous NIV systems remain uncommon, the review adopts a pragmatic definition: technologies are considered relevant when they use physiological or ventilator-derived feedback to modify ventilatory assistance, trigger/cycling behavior, delivered pressure, volume targeting, or clinical decision support. In control terms, a closed loop requires a reference or target, a measured or estimated output that is fed back to the controller, and an actuator that changes ventilatory assistance (9). Monitoring variables are therefore feedback elements, whereas clinical context and clinician-defined goals provide supervisory inputs. The central argument is that current NIV practice already contains several partial loops, but these components have not yet been integrated into robust, outcome-tested closed-loop NIV platforms. Figure 1 summarizes this architecture and its safety constraints.

What should count as closed-loop NIV?

Closed-loop ventilation is best understood as a control architecture rather than as a single proprietary ventilator mode. Chatburn and Mireles-Cabodevila classified ventilator targeting schemes as set-point, dual, servo, adaptive, optimal, and intelligent according to how preset goals are achieved (10). Lellouche and Brochard emphasized that PAV, NAVA, ASV, and SmartCare represent advanced closed loops with different inputs, response times, and clinical objectives (11). Van der Staay and Chatburn further showed that “optimal” targeting depends on the objective function: at comparable simulated ventilation efficiency in adult ARDS, expected tidal volume was approximately 7.4 mL/kg with ASV, 6.2 mL/kg with AVM2, and 6.7 mL/kg with MFV, with different optimal frequencies generated by the different cost functions (12). Thus, automation embeds a selected physiological objective rather than eliminating clinical judgment.

For NIV, this taxonomy matters because the term “closed-loop NIV” can be overused. AVAPS and iVAPS adapt inspiratory pressure to a ventilation target but do not monitor effort or failure risk; leak-compensation algorithms modify triggering and flow estimation without choosing clinical goals; and NAVA provides servo assistance proportional to EAdi without deciding whether NIV remains appropriate. A mature system would probably be hierarchical and multirate: fast breath-to-breath loops would address leak, triggering, and cycling, whereas slower loops would target ventilation, effort, gas exchange, and failure risk. Stability requires accounting for sensor latency, physiological delays, signal noise, inter- and intra-patient variability, competing targets, and nonlinear or path-dependent behavior. In the respiratory system, pressure-volume hysteresis can reflect recruitment/derecruitment and viscoelastic behavior, illustrating that the response to a given pressure may depend on the preceding state (13). In physiological closed-loop devices, sensor latency and dynamic uncertainty can impair controller stability, supporting bounded rates of change, fail-safe or fallback modes, and clinician override (9, 14). Feedback inputs should distinguish directly measured device signals (airway pressure, device flow), model-derived estimates (leak, patient flow, tidal volume, mechanics), and additional physiological measurements or surrogates (EAdi, esophageal pressure, P0.1, SpO2, ultrasound), which differ in measurement pathway, susceptibility to artifact or estimation error, and temporal resolution (14–17). These categories should not be interpreted as an ordinal reliability scale; signal validity is device-, interface-, and context-dependent.

The evidence base in invasive ventilation shows both the potential and the limits of this concept. A systematic review found that closed-loop ventilation may improve process measures and reduce staff workload, although the effects on clinical outcomes vary across systems and populations (18). The large ACTiVE trial reported that early INTELLiVENT-ASV did not increase ventilator-free days compared with protocolized conventional invasive ventilation, despite improved ventilation quality and fewer severe gas-exchange derangements (19). These findings are relevant to NIV because they caution against assuming that better automation will necessarily translate into patient-centered benefit.

Leak compensation, triggering, and cycling: the first practical NIV loops

The most concrete closed-loop problem in NIV is the leak. In a bench model, Vignaux and colleagues demonstrated that leaks impair ICU ventilator functions by increasing trigger delay and triggering workload, reducing pressurization, and delaying cycling; NIV-specific modes partially corrected these effects, but performance varied widely, and some ventilators worsened leak-related dysfunction (3). In a multicenter clinical study, severe asynchrony during NIV for acute respiratory failure was common: an asynchrony index greater than 10% occurred in 43% of patients, and leak magnitude correlated with ineffective efforts and delayed cycling (20). A subsequent randomized crossover clinical study showed that NIV algorithms reduced leak-related asynchronies, but global asynchrony did not uniformly improve because some algorithms increased premature cycling (21). These studies define a core principle for closed-loop NIV: an algorithm that corrects one form of asynchrony can create another if it is not physiologically integrated.

Carteaux and colleagues extended this point by comparing ICU, transport, and dedicated NIV ventilators in bench and clinical settings. Dedicated NIV ventilators achieved lower asynchrony indices than ICU ventilators with or without their NIV algorithms, largely because of better handling of leaks and auto-triggering (4). Recent bench evidence and a 2026 review of patient-ventilator interaction confirm that modern devices still vary in leak handling and that unintentional leak remains a major determinant of trigger and cycle discordance (5, 22, 23). Trigger and cycling asynchronies, including ineffective efforts, auto-triggering, and delayed or premature cycling, should therefore be interpreted together with leak magnitude, pressure/flow waveform morphology, respiratory mechanics, and support level rather than by a global asynchrony index alone (20–25). Thus, accurate estimation of patient flow despite intentional and unintentional leaks remains a prerequisite for advanced NIV control (15).

Cycling-off is particularly important. Calderini and colleagues showed that, in the presence of mask leaks, time-cycled inspiratory pressure support improved patient-machine interaction compared with conventional flow-cycled NIV in severe hypoxemic respiratory failure (26). Tassaux and colleagues later showed in intubated patients with obstructive lung disease that increasing the expiratory trigger threshold reduced delayed cycling, intrinsic PEEP, non-triggering breaths, and inspiratory muscle workload (27). Although this latter study was not performed during NIV, its physiological mechanism is directly relevant to NIV algorithms in COPD and other obstructive conditions. Chen and colleagues tested automated cycling during noninvasive pressure support ventilation in a lung simulator and found better synchronization with Auto-Trak than with conventional cycling approaches across simulated mechanics and leaks (28).

Waveform-guided synchronization and proportional assistance

A key step toward clinically useful closed-loop NIV is the ability to infer the timing of patient effort from ventilator waveforms without invasive monitoring. Mojoli and colleagues developed a waveform method using airway pressure and flow, with esophageal pressure as the reference, that accurately detected respiratory efforts, trigger delay, cycling delay, and early cycling (24). The same group then tested automated real-time waveform-guided cycling-off in 15 difficult-to-wean patients. Automation reduced cycling delay, ineffective efforts, and total asynchrony time compared with standard cycling. It also performed as well as or better than protocol-defined expert adjustment of expiratory trigger sensitivity, but the restricted expert intervention precludes interpreting this as superiority over comprehensive expert ventilator management (25). Although conducted during invasive pressure support, these studies provide a mechanistic bridge to NIV, where continuous noninvasive effort-timing detection would be particularly valuable.

Direct NIV evidence is also emerging. Renaud and colleagues evaluated an automated real-time waveform analysis algorithm during NIV in a bench model and reported improved patient-ventilator synchrony under leak conditions (22). This study is highly relevant because it tests a technological solution tailored to variable leaks that distort pressure and flow signals. However, it used a single ventilator, calibrated circuit leaks without a face-mask interface, and three predefined simulated respiratory-mechanics conditions; therefore, the evidence remains preclinical and cannot establish performance across devices or clinical NIV conditions. Clinical trials are still needed to determine whether automated waveform-guided NIV reduces discomfort, intubation, sedation use, or time under ineffective support.

NAVA represents a different strategy: instead of inferring neural timing from pneumatic waveforms, it uses EAdi to trigger, cycle, and proportionally assist each breath. Piquilloud and colleagues showed that NAVA improved patient-ventilator interaction during NIV delivered by face mask compared with conventional pressure support (29). Longhini and colleagues later showed that optimized NIV-NAVA settings could improve comfort and synchrony compared with pneumatic triggering and cycling (30). Randomized clinical evidence tempers these physiological findings: in adults with de novo acute respiratory failure, NAVA reduced asynchrony and NIV-related complications but not NIV failure or 28-day mortality compared with pressure support ventilation (31). In AECOPD, NAVA and ASV had similar NIV failure, asynchrony, and 90-day mortality, although ASV required fewer clinician adjustments (32). A 2026 feasibility trial of ASV with IntelliSync versus pressure support ventilation in AECOPD found greater comfort and shorter hospital stay, with a numerically lower but nonsignificant NIV-failure rate; larger trials are required (33). PAV and PAV + remain conceptually relevant because they adjust support in proportion to patient effort and mechanics, but evidence in critical care NIV is less developed than in invasive assisted ventilation (11).

Volume-assured and adaptive pressure modes

Volume-assured pressure support modes are not full closed-loop NIV systems, but they represent a clinically available form of adaptive NIV. AVAPS adjusts inspiratory pressure support within a prescribed range to maintain a target tidal volume, while iVAPS targets alveolar ventilation and may incorporate auto-EPAP in selected devices. Shah and colleagues summarized these modes as extensions of bilevel NIV intended to stabilize ventilation when mechanics or effort change over time (34). In acute hypercapnic respiratory failure, Briones Claudett and colleagues reported clinical experience with BiPAP S/T-AVAPS, supporting the plausibility of adaptive pressure targeting in the acute setting (35). In chronic hypercapnic COPD, AVAPS and iVAPS studies suggest potential benefits in gas exchange, symptoms, and quality of life, although populations, devices, and treatment settings differ substantially (36, 37).

The most recent synthesis on this topic is the 2026 systematic review and meta-analysis by Gregory and colleagues. Across 11 comparative studies involving adults with acute hypercapnic respiratory failure or similar acute-care indications, AVAPS was associated with a lower risk of intubation than BiPAP, without a clear short-term mortality benefit and with similar physiological outcomes and hospital length of stay (38). However, only three studies contributed complete intubation data; the evidence mixed randomized and observational designs, was dominated by COPD populations, used variable outcome time points, and carried low-to-moderate risk-of-bias concerns. The pooled finding should therefore be interpreted as evidence for AVAPS in selected acute hypercapnic populations rather than evidence that closed-loop NIV as a class improves clinical outcomes. AVAPS remains a pressure-adaptation loop targeting volume and does not directly measure patient effort, dyspnea, synchrony, or progression toward failure.

Effort, drive, and failure prediction as missing loop inputs

A major limitation of current NIV automation is that most systems respond to pressure, flow, volume, and leaks, but not to the patient's injurious effort or probability of NIV failure. This gap is clinically important. Noninvasive respiratory support can reduce the work of breathing, but persistent high respiratory drive and effort may promote respiratory muscle fatigue, cardiovascular compromise, oxygen-delivery imbalance, and potentially patient self-inflicted lung injury (6, 8). The same review literature emphasizes that delayed intubation remains a central mechanism of harm when noninvasive support fails (2, 6).

Inspiratory effort monitoring may therefore become a key input for closed-loop NIV. Tonelli and colleagues studied patients with de novo hypoxemic respiratory failure undergoing NIV and found that relief of inspiratory effort, assessed by esophageal pressure within the first 2 h, was an early and accurate predictor of NIV outcome (7). This study is especially relevant because it connects physiological monitoring to a clinically actionable decision: whether NIV is unloading the patient sufficiently or merely postponing intubation. Telias and Spadaro reviewed methods to monitor respiratory drive and effort, including esophageal pressure, airway occlusion pressure, EAdi, and ultrasound, while emphasizing the need to interpret these signals in relation to respiratory muscle function and ventilator settings (16). Cammarota and colleagues further positioned esophageal pressure, EAdi, electrical impedance tomography, lung ultrasound, and respiratory muscle ultrasound as advanced bedside tools for personalized protection of the lung and respiratory muscles during acute respiratory failure (17).

Future NIV controllers should therefore move from synchrony loops toward injury-avoidance loops. Such systems would not simply minimize trigger delay or achieve a target tidal volume. They would aim to maintain adequate unloading while avoiding excessive tidal volume, excessive negative pleural pressure swings, high respiratory rate, worsening gas exchange, and poor comfort. The clinical output may sometimes not be a ventilator adjustment, but a recommendation to change interface, increase monitoring intensity, or proceed to intubation.

Artificial intelligence and future closed-loop NIV

Artificial intelligence (AI) may help integrate heterogeneous data streams that are difficult to combine through rule-based control alone. However, model performance depends on how variables and time are represented. Across ventilator-waveform and electronic health record applications, feature selection, sampling frequency, aggregation windows, temporal alignment, prediction horizon, missing-data handling, label definition, and class balance determine what physiology a model can learn; a breath-by-breath signal and a 30-minute trend are not interchangeable inputs. A systematic methodological review of longitudinal electronic health record prediction models highlights the importance of these design choices for temporally structured clinical data (39). Current AI work in ventilation remains largely retrospective, simulation-based, or invasive rather than NIV-specific, so the near-term opportunities are signal interpretation, phenotyping, failure prediction, and safety testing rather than autonomous control.

Three lines of work are particularly relevant. Ontology-based representations may standardize NIV concepts across electronic health records, devices, indications, parameters, and outcomes; Islam, Mosier, and Subbian developed and evaluated such an ontology for acute care to support interoperability, rule-based reasoning, and future decision support (40). Computational leak estimation can improve patient-flow reconstruction, a prerequisite for accurate waveform interpretation and automated cycling (15). Finally, reinforcement learning and digital twins may permit safety testing of ventilator policies before clinical exposure. Yu and colleagues proposed a safety-oriented offline reinforcement learning framework with temporal modeling and digital-twin verification for automated mechanical ventilation (41). Although not NIV-specific and currently a preprint, it illustrates why future algorithms must be uncertainty-aware, auditable, and validated under physiologically plausible scenarios.

For critical care NIV, AI-enabled closed-loop systems should be designed as supervised clinical tools rather than autonomous substitutes for clinicians. The clinician should remain in the loop, especially because NIV failure is a clinical diagnosis influenced by gas exchange, work of breathing, mental status, hemodynamics, secretion burden, airway protection, patient goals, and local expertise. AI may identify patterns that humans miss, but it must also be auditable, calibrated across populations, robust to missing data, and constrained by safety rules.

Discussion

Closed-loop NIV is best viewed as an emerging continuum rather than a mature therapeutic class. The most established components are leak-adapted triggering/cycling, dedicated NIV algorithms, and proportional or neural assistance. Direct NIV studies support improved synchrony and, in selected studies, comfort or fewer ventilator adjustments, but randomized NAVA/ASV evidence has not shown a consistent reduction in NIV failure, and the recent ASV-IntelliSync result remains feasibility evidence (4, 20–22, 29–33). Adaptive pressure modes such as AVAPS and iVAPS extend the concept by targeting ventilation, but outcome evidence remains uncertain and phenotype-dependent (38). In our view, the near-term role of closed-loop NIV should be bounded automation: machines can optimize rapidly changing technical tasks such as leak compensation, triggering, cycling, and proportional assistance, while clinicians retain control of therapeutic goals, escalation, and the decision to intubate. Table 1 summarizes these strategies, their feedback signals, automated outputs, clinical evidence, and main limitations.

This distinction is clinically important because NIV failure is not a single control error. Overreliance on automation may mask deterioration or delay intubation, whereas well-designed automation may reduce workload, practice heterogeneity, and delayed adjustment of settings. The invasive ventilation literature illustrates this balance: closed-loop systems can improve process quality without consistently improving patient-centered outcomes (18, 19). We therefore do not consider current evidence sufficient for an autonomous algorithm to determine continuation of NIV. Future studies should instead evaluate defined closed-loop functions in clinically coherent phenotypes, including hypercapnic COPD, obesity hypoventilation, cardiogenic pulmonary edema, post-extubation and postoperative support, and carefully selected hypoxemic respiratory failure.

Clinically relevant endpoints should include comfort, dyspnea, inspiratory effort relief, asynchrony burden, delivered tidal volume, gas-exchange stability, staff workload, intubation, delayed intubation, and patient-centered outcomes. A particularly informative trial would compare standard NIV with a supervised hierarchical strategy combining leak compensation, automated cycling, effort monitoring, and explicit failure-alert criteria. The system should disclose which feedback signal triggered an adjustment or alert and operate within predefined safety limits, allowing the clinician to distinguish technical optimization from a recommendation for escalation.

In summary, current evidence supports the physiological rationale for closed-loop NIV but not autonomous management or a major clinical-outcome benefit. We propose a hierarchical, multirate architecture in which fast technical loops manage leak and synchrony, intermediate loops regulate ventilation and unloading, and a slower supervisory layer integrates gas exchange, comfort, clinical trajectory, and failure risk. Such systems should remain clinician-supervised, transparent, and safety-constrained until prospective trials demonstrate that automation improves outcomes without delaying escalation.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this manuscript, FG used OpenAI ChatGPT, GPT-5.5 Thinking model, OpenAI, to assist with literature organization, drafting support, language editing, reference checking, and preparation of the conceptual figure. OpenAI image-generation tools available within ChatGPT were also used to assist in creating Figure 1. The authors reviewed, edited, and verified all AI-assisted text, citations, references, and visual content, ensured that the figure accurately reflects the concepts discussed in the manuscript, and take full responsibility for the final content of the article. No generative AI tool is listed as an author.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Marco Fiore, Università degli Studi della Campania 'Luigi Vanvitelli', Italy

Reviewed by: Ivaylo Minev, Plovdiv Medical University, Bulgaria