Physiology-guided mechanical ventilation: Monitoring, proportional assist, and bounded automation

Abstract

Mechanical ventilation has evolved into a complex intervention that influences lung injuries, respiratory muscle function, and hemodynamic stability. Although lung-protective strategies improve outcomes in acute respiratory distress syndrome, bedside management remains limited by incomplete monitoring of key physiologic variables, including lung stress, inspiratory effort and regional ventilation. This constrains decision such as positive end-expiratory pressure titration and ventilatory assist targeting. Emerging technologies aim to address these gaps by improving physiological assessment and enabling more individualized care. Tools such as esophageal manometry, airway occlusion pressure (P0.1), diaphragm electrical activity, and electrical impedance tomography provide insight into lung mechanics, respiratory drive, and regional ventilation. Proportional modes of ventilation improve patient-synchrony, though their impact on patient-centered outcomes remains variables. Automation and artificial intelligence are increasingly applied to ventilator management, supporting wave analysis, detection of asynchrony, and prediction of weaning readiness. These tools may also assist clinical decision-making within predefined safety limit. We propose a pragmatic, clinical-directed framework integrating physiologic monitoring, proportional assist, and bounded decision support to optimize lung protection, diaphragm function, and hemodynamica stability.

Key Words: Mechanical ventilation; Critical care; Pulmonary medicine; Intensive care medicine; Ventilator management

Core Tip: Mechanical ventilation should be approached as a dynamic, patient-specific therapy rather than a fixed protocol. Integrating physiologic monitoring of respiratory effort, lung stress, and regional ventilation can help clinicians individualize support while balancing lung protection and diaphragm preservation. For example, ventilatory assist may be titrated to maintain moderate inspiratory effort rather than targeting airway pressures or tidal volume alone. Emerging monitoring, proportional assist modes, and bounded automation are best viewed as tools that augment physiologic assessment while preserving clinician judgment in the management of critically ill patients.

INTRODUCTION

Mechanical ventilation is a cornerstone therapy in critical care, yet it remains one of the most consequential interventions delivered at the bedside[1-4]. While modern ventilators reliably support oxygenation and carbon dioxide elimination, contemporary goals extend beyond gas exchange to include limitation of ventilator-induced lung injury (VILI), preservation of respiratory muscle function, optimization of patient-ventilator interaction, and maintenance of hemodynamic stability[3,4]. These objectives evolve dynamically as lung mechanics, respiratory drive, and cardiovascular reserve change over time.

Mechanical ventilation is increasingly conceptualized as a form of precision dosing. In this framework, the delivered “dose” includes not only tidal volume and airway pressure, but also flow pattern, timing, spontaneous effort, and the regional distribution of stress and strain within a heterogeneous lung[4,5]. For example, two patients receiving identical tidal volumes may experience markedly different transpulmonary pressures and lung stress due to differences in chest wall elastance, abdominal pressure, or patient effort. This highlights the limitations of relying on airway pressure or tidal volume alone. In this framework, “toxicities” include excessive transpulmonary pressure, injurious inspiratory effort, and prolonged diaphragm inactivity, whereas the therapeutic window lies in achieving adequate gas exchange while maintaining lung stress and respiratory effort within physiologic bounds[6,7].

Despite this conceptual model, bedside titration remains largely guided by global surrogates such as airway pressures and gas exchange. Key physiologic domains, including lung stress, respiratory effort, and regional ventilation, are often assessed indirectly and in isolation. This fragmented approach limits the ability to understand how these variables interact and impairs individualized ventilator adjustment[8-12]. Because no single measurement captures these processes, individualized management requires integration of complementary physiologic signals rather than reliance on any single parameter.

Prior reviews have examined monitoring tools, ventilatory modes, and computational approaches largely in isolation, without providing a cohesive framework for integrating these signals into real-time bedside decision-making. In contrast, this review proposes an integrated, physiology-guided framework that unifies respiratory effort, lung mechanics, regional ventilation, and cardiopulmonary interactions into a coordinated clinical strategy for ventilator titration. This framework emphasizes the combined use of advanced monitoring tools, proportional modes of ventilation, and bounded decision-support systems to guide dynamic adjustment of ventilatory support[13-17].

Despite strong physiologic rationale, outcome-driven evidence supporting physiology-guided ventilation strategies remains limited, and the clinical impact of many advanced monitoring tools and proportional modes continues to be actively investigated.

LITERATURE REVIEW

We performed a narrative review of the literature using PubMed, Google Scholar, and the Cochrane Library (last search: December 2025), focusing on intensive care unit (ICU)-based studies of invasive mechanical ventilation in adults (including mixed adult/pediatric cohorts when highly relevant). We targeted literature related to: (1) Mechanistic understanding of VILI and ventilator-induced diaphragm dysfunction (VIDD); (2) Patient-ventilator interaction and proportional ventilatory assistance; (3) Advanced monitoring of lung stress/strain, respiratory effort, and regional ventilation; and (4) Automation and artificial intelligence (AI)/machine-learning (ML) applications in invasive mechanical ventilation. When available, we prioritized higher-level clinical evidence (guidelines, randomized trials, and meta-analyses), followed by observational studies and foundational physiologic/mechanistic literature. Search terms included combinations of “mechanical ventilation”, “patient-ventilator asynchrony”, “respiratory drive”, “P0.1”, “esophageal pressure”, “electrical impedance tomography”, “neurally adjusted ventilatory assist”, “proportional assist ventilation”, “closed-loop ventilation”, and “artificial intelligence”, with representative search strings such as: “Mechanical ventilation” AND “respiratory drive” OR “P0.1” OR “EAdi”, (“mechanical ventilation” AND “esophageal pressure” OR “transpulmonary pressure”), and (“mechanical ventilation” AND “electrical impedance tomography” OR “EIT”), applying filters for human studies and adult ICU populations when appropriate. We preferentially reviewed publications from the past decade but included foundational physiologic and clinical studies when needed for context. Across these domains, approximately 300 titles and abstracts were screened, with approximately 150 full-text articles reviewed. Studies were included if they addressed invasive mechanical ventilation in ICU settings and provided mechanistic, physiologic, or clinical outcome data relevant to ventilator management, with exclusion of studies focused exclusively on non-invasive ventilation, chronic outpatient ventilation, or non-clinical experimental models without clear physiologic or translational relevance. Studies were synthesized using a thematic framework based on key physiologic domains, including respiratory effort, lung mechanics and stress/strain, regional ventilation, patient-ventilator interaction, and automation/AI-supported decision-making. When evidence was conflicting, we summarized both positive and negative trials and prioritized higher-level evidence (guidelines, randomized trials, and meta-analyses) without excluding studies based on direction of findings. When differences in study interpretation arose, these were resolved through discussion among co-authors with senior author oversight to ensure consistency in physiologic interpretation and clinical framing. When mechanistic and clinical outcome data were discordant, greater weight was given to higher-level clinical evidence when available, while mechanistic studies were used to contextualize physiologic plausibility and inform interpretation in areas where clinical data remain limited. This approach aligns with key domains of high-quality narrative reviews as outlined by the Scale for the Assessment of Narrative Review Articles, including clearly defined objectives, structured literature search, transparent reporting of sources, balanced presentation of evidence, and integration of mechanistic and clinical perspectives. Selection bias is an inherent limitation of narrative reviews. Inclusion of mixed adult/pediatric studies was based on physiologic relevance, which may introduce heterogeneity in applicability to strictly adult ICU populations. This was not a systematic review and formal risk-of-bias assessment was not performed. No individual patient data were used in this study.

NEURAL RESPIRATORY DRIVE AS A CONTROL SIGNAL FOR ASSISTED VENTILATION

Why does monitoring inspiratory effort matter at the bedside?

Ventilation is regulated through integrated brainstem networks that balance metabolic demand with ventilatory output. This makes assessment of neural respiratory drive central to understanding patient effort during assisted ventilation. In adults, PaCO₂/pH is the dominant physiologic driver of minute ventilation across many conditions, sensed primarily via central chemoreception, while hypoxemia is sensed mainly via peripheral carotid chemoreceptors and becomes a stronger driver as PaCO₂ falls[18-20]. Pulmonary mechanoreceptors and vagal afferents modulate breathing pattern and reflex responses (e.g., changes in inspiratory time or tachypnea) to mechanical stimuli, but these reflexes do not reliably “protect” mechanically ventilated patients from overdistension or barotrauma[21-23].

When does spontaneous breathing become injurious rather than protective?

In critical illness, neural drive is frequently perturbed by pain, anxiety, fever, acid-base abnormalities, hypoxemia/hypercapnia, lung inflammation, and sedative or opioid exposure. The resulting spectrum ranges from absent or blunted drive (deep sedation, encephalopathy, neuromuscular blockade) to excessive drive (metabolic acidosis, severe dyspnea, patient distress)[24]. This matters because patient effort interacts with ventilator settings. Together, these determine the achieved transpulmonary pressure, tidal volume, and flow profile. Clinically, respiratory effort can be conceptualized along a continuum from insufficient effort, which may promote diaphragm unloading and atrophy, to excessive effort associated with large negative pleural pressure swings and increased regional lung stress; however, precise physiologic thresholds defining these states remain uncertain and likely vary across disease states and stages of illness. When effort is excessive or poorly supported, large negative pleural pressure swings can raise transpulmonary pressure and regional lung stress despite apparently “protective” airway pressures, contributing to patient self-inflicted lung injury (P-SILI)[6,7].

Signals that reflect neural drive [e.g., diaphragm electrical activity (EAdi)] are therefore attractive targets for assisted ventilation and monitoring. In principle, coupling ventilator triggering, cycling, and assist magnitude to neural output can reduce dyssynchrony and better match delivered support to patient demand across changing clinical states[13-15]. Similarly, EAdi provides continuous insight into neural drive, where persistently elevated values may indicate under-assistance or high respiratory demand, while very low or absent activity may reflect over-assistance, excessive sedation, or neuromuscular dysfunction[13-15]. At the bedside, airway occlusion pressure (P0.1) is often used as a surrogate of respiratory drive, with higher values suggesting increased respiratory drive and potential risk of injurious effort, although interpretation remains context dependent. Importantly, no single effort metric has been validated as an outcome-driven target. Ongoing research aims to better define clinically actionable thresholds and composite approaches. In practice, no single effort metric should guide ventilator titration in isolation. A pragmatic bedside approach may integrate P0.1, EAdi trends when available, ventilator waveform assessment, tidal volume variability, and clinical evaluation of patient comfort or distress to estimate whether inspiratory effort is within a physiologically acceptable range.

NEURAL RESPIRATORY DRIVE AS A CONTROL SIGNAL FOR ASSISTED VENTILATION

When does spontaneous breathing become injurious rather than protective?

The mechanical load imposed on the lung is best conceptualized as stress (pressure applied to lung tissue, approximated by transpulmonary pressure) and strain (deformation relative to a reference lung volume). At the bedside, clinicians commonly rely on airway pressure and global compliance measurements, but these variables represent the combined respiratory system (lung + chest wall) and do not capture regional heterogeneity or the patient’s contribution to the breath[25-29]. Importantly, compliance and elastance are global properties that can change rapidly with recruitment, derecruitment, fluid balance, and body position.

In acute respiratory distress syndrome (ARDS) and other forms of acute lung injury, the “baby lung” concept highlights that functional aerated lung volume is often reduced; thus, a fixed tidal volume (even when indexed to predicted body weight) may impose disproportionate strain on the remaining aerated units[27-29]. At the microscopic scale, cyclic opening and closing of unstable units can generate high local stress concentrations “atelectrauma”, while overdistension of relatively compliant regions can occur simultaneously in other lung areas[5,26,27,30]. These mechanical insults contribute to endothelial and epithelial injury and can amplify systemic inflammation “biotrauma” through release of proinflammatory mediators such as IL-6, IL-8, and TNF-α[28,29].

How should clinicians interpret measures of ventilator intensity?

Driving pressure and mechanical power have emerged as integrative risk markers that incorporate aspects of tidal strain and the frequency-dependent energy delivered to the respiratory system[10,11]. While such metrics are conceptually useful, their interpretation depends on lung size, recruitability, and chest wall mechanics; therefore, they should be viewed as complementary-rather than definitive-targets that may benefit from adjunctive monitoring of transpulmonary pressure and regional ventilation distribution[8-12].

Taken together, driving pressure, mechanical power, and transpulmonary pressure reflect related but distinct aspects of ventilator intensity and lung stress and should not be used interchangeably. Driving pressure reflects tidal volume relative to global respiratory system compliance and is most interpretable when chest wall mechanics are stable[10,11]. When chest wall elastance is increased, airway pressures may overestimate true lung stress[8-10]. Mechanical power integrates pressure, volume, flow, and respiratory rate to estimate energy delivery per unit time and may capture cumulative risk not apparent from tidal volume or plateau pressure alone, but it remains primarily a risk marker rather than a validated bedside titration target[11,31]. Transpulmonary pressure, estimated via esophageal manometry, more directly approximates lung-distending stress when lung and chest wall contributions diverge, though it is technique-dependent and must be interpreted within clinical context[8-10,16]. For example, in patients with obesity, abdominal hypertension, or chest wall edema, elevated pleural pressures may result in high airway plateau pressures that overestimate true lung stress, making transpulmonary pressure a more informative parameter in these settings[8-10].

How should these metrics be applied at the bedside?

In practice, these measures are complementary rather than hierarchical. Driving pressure and mechanical power provide global risk context, whereas transpulmonary pressure may offer additional physiologic precision in selected patients, particularly when chest wall mechanics are altered. No single metric captures regional heterogeneity or patient effort in isolation; therefore, ventilator management should integrate these parameters within a broader multimodal clinical framework.

DIAPHRAGM LOAD: AVOIDING BOTH DISUSE AND INJURIOUS EFFORT

How can clinicians preserve diaphragm function without compromising lung protection?

The diaphragm is the primary inspiratory muscle and supports ventilation through generation of negative pleural pressure. It also contributes to venous and lymphatic return and modulates cardiac loading conditions through changes in intrathoracic pressure[32-34]. Mechanical ventilation can disrupt this physiology in two opposing ways. First, controlled ventilation with suppressed drive can produce diaphragmatic inactivity, promoting VIDD through disuse atrophy and reduced contractile efficiency. Second, during assisted ventilation, insufficient ventilatory support or high respiratory drive can lead to excessive inspiratory effort and potentially injurious diaphragm loading[6,7]. The clinical goal, therefore, is not simply to restore spontaneous breathing, but to titrate ventilatory assistance to maintain sufficient inspiratory effort to preserve diaphragm activity and neuromechanical coupling, while avoiding excessive effort that may produce large pleural pressure swings, dynamic hyperinflation, or substantial increases in transpulmonary pressure, although precise physiologic thresholds defining this range remain uncertain[6,7]. This concept underpins lung- and diaphragm-protective ventilation, which explicitly integrates lung stress/strain targets with monitoring and management of respiratory drive/effort[6,7]. However, continuous monitoring of respiratory effort using advanced modalities such as EAdi or esophageal pressure is not universally available, and clinicians often rely on indirect measures including ventilator waveforms and clinical assessment in typical ICU settings.

If inspiratory effort remains high despite assisted or proportional ventilation, first correct reversible causes such as pain, anxiety, fever, hypoxemia, metabolic acidosis, and dyssynchrony, and reassess ventilator settings. Persistent excessive effort may indicate that the patient’s respiratory load exceeds what can be safely supported within lung-protective limits, and temporary use of controlled ventilation or short-term deeper sedation may be necessary to prevent injurious transpulmonary pressures. Conversely, brief reductions in diaphragm activity may be acceptable early in severe lung injury when strict lung protection is required, but prolonged suppression increases the risk of VIDD and delayed weaning. These considerations are particularly relevant during weaning, where both excessive inspiratory effort and diaphragm weakness may contribute to weaning failure[6,7]. In practice, management requires ongoing reassessment and balancing lung protection with preservation of diaphragm function[6,7].

CARDIOPULMONARY INTERACTIONS DURING POSITIVE-PRESSURE VENTILATION

How do ventilator settings influence right ventricular function and hemodynamics?

Positive-pressure ventilation alters intrathoracic pressures and can meaningfully affect cardiac preload, afterload, and right ventricular (RV) performance[35-38]. Increases in pleural pressure reduce the gradient for venous return and may decrease right atrial filling and RV preload; simultaneously, increases in alveolar pressure (particularly at high lung volumes) can raise pulmonary vascular resistance and RV afterload[35-38]. The net effect is patient-specific and depends on volume status, vasomotor tone, ventricular function, and lung mechanics.

Clinically, these interactions are most evident during initiation of positive-pressure ventilation, escalation of positive end-expiratory pressure (PEEP), or recruitment maneuvers in patients with shock, RV dysfunction, or pulmonary hypertension. A rise in RV afterload can lead to RV dilation and interventricular septal shift, impairing left ventricular filling and systemic output[38]. At the bedside, additional warning signs of adverse cardiopulmonary interactions may include rising central venous pressure and new or worsening hypotension following increases in PEEP, which may indicate impaired venous return or worsening RV loading conditions[35-38]. Conversely, in selected patients with left ventricular systolic dysfunction, modest increases in pleural pressure can reduce Left Ventricular transmural pressure and afterload, improving cardiac output[36,37].

Because pleural pressure and transpulmonary pressure are distinct P_L= P_alv- P_plstrategies designed to limit lung stress (transpulmonary pressure) should not be conflated with hemodynamic effects driven predominantly by pleural pressure. This distinction is particularly relevant when interpreting airway pressures in patients with altered chest wall mechanics or increased intra-abdominal pressure, where a higher airway plateau pressure may not imply proportionally higher transpulmonary pressure[8-10].

These cardiopulmonary interactions have direct implications for ventilator strategy. Escalation of PEEP or recruitment maneuvers may improve oxygenation but can simultaneously increase pulmonary vascular resistance and RV afterload, particularly in patients with shock, pulmonary hypertension, or pre-existing RV dysfunction[35-38]. Excessive lung distention may further impair RV output by increasing alveolar pressures and compressing pulmonary capillaries[35-38]. Conversely, spontaneous inspiratory effort generates negative pleural pressure swings that augment venous return and RV preload[35-38]. Large swings may also increase transpulmonary pressure and RV wall stress when effort is excessive[6-10].

Accordingly, decisions regarding PEEP titration, permissive spontaneous breathing, and recruitment strategies should incorporate assessment of RV function and hemodynamic response rather than relying solely on oxygenation or airway pressure targets. Interpreting airway and plateau pressures without considering chest wall mechanics and pleural pressure may obscure their true cardiopulmonary impact[8-10]. These findings should prompt reassessment of ventilator strategy, including reduction of excessive PEEP, adjustment of tidal volume or driving pressure, and evaluation of volume status and RV function. Integrating these considerations reinforces the broader framework of precision ventilation, in which lung stress, respiratory effort, and RV performance are evaluated together rather than in isolation.

PATIENT-VENTILATOR ASYNCHRONY AND THE LIMITATIONS OF DEEP SEDATION/NEUROMUSCULAR BLOCKADE

Patient-ventilator asynchrony (PVA) refers to discordance between patient effort and ventilator delivery. Common forms include ineffective triggering, double triggering/breath stacking, flow mismatch (flow starvation or excessive flow), and cycling asynchrony (premature or delayed cycling)[39-42]. PVA can increase work of breathing, worsen comfort, and prompt escalation of sedation. In observational studies, clinically significant asynchrony is associated with longer duration of ventilation and ICU stay, though causality is difficult to establish because dyssynchrony, sedation depth, and illness severity are tightly linked[39-42]. Importantly, the relationship between patient–ventilator asynchrony and adverse outcomes is associative, and causality remains difficult to establish given confounding factors such as disease severity, sedation practices, and underlying respiratory mechanics.

Some patterns have direct physiologic implications. Double triggering can increase delivered tidal volume and end-inspiratory lung stress, potentially undermining lung-protective targets. Severe flow starvation may increase inspiratory effort and pleural pressure swings, increasing transpulmonary pressure and P-SILI risk, whereas excessive flow and inappropriate cycling can contribute to dynamic hyperinflation and patient discomfort[40-42].

Deep sedation and neuromuscular blockade can suppress respiratory drive and reduce some types of dyssynchrony, but they carry well-described harms including delirium, immobility, ICU-acquired weakness, and prolonged rehabilitation[43,44]. Neuromuscular blockade may improve synchrony and facilitate strict lung-protective ventilation and prone positioning in early severe ARDS, yet evidence is mixed and guideline recommendations are conditional; routine prolonged paralysis is not supported[7,45]. Therefore, a practical hierarchy is to first optimize ventilator settings and mode (trigger sensitivity, cycling criteria, inspiratory flow delivery, PEEP, and level of assist) and treat reversible drivers of high drive (pain, anxiety, fever, acidosis), reserving deeper sedation or short courses of paralysis for refractory hypoxemia, proning facilitation, or persistent injurious dyssynchrony despite ventilator optimization[7,45].

In practice, management of asynchrony should begin with identification of the predominant asynchrony pattern, followed by targeted adjustment of ventilator settings or sedation strategy. For example, trigger asynchrony may improve with adjustment of sensitivity or reduction of intrinsic PEEP, whereas double triggering may require reduction in respiratory drive or adjustment of inspiratory time and tidal volume.

CHALLENGES IN CONDITIONS WITH RAPIDLY EVOLVING PHYSIOLOGY

These exhibit substantial inter- and intra-patient variability in recruitability, morphology, including focal vs nonfocal patterns, and chest wall mechanics[46-50]. Cardiopulmonary reserve and overall physiology evolve over time, such that ventilator settings appropriate at one stage may become harmful as the disease progresses or resolves[45,48-51]. Additionally, coexisting shock and fluid resuscitation can worsen pulmonary edema and lung mechanics, while higher airway pressures and PEEP can destabilize RV function and systemic perfusion[52,53].

In this context, clinicians must continuously balance competing objectives: Oxygenation and CO₂ clearance; limitation of lung stress/strain; preservation of diaphragm function; minimization of oversedation; and maintenance of adequate perfusion. This multi-objective problem motivates the emergence of monitoring and automation strategies that quantify key physiologic states (effort, stress, recruitment, hemodynamics) and support timely adaptation of ventilator settings to changing conditions.

For clarity, the technologies discussed in this review span a spectrum of clinical readiness. First, established physiologic tools with defined bedside roles (e.g., esophageal pressure monitoring) are supported by mechanistic validation and selected clinical trials. Second, emerging but increasingly actionable approaches [e.g., proportional modes of ventilation and regional ventilation assessment with electrical impedance tomography (EIT)] demonstrate consistent physiologic benefits, although patient-centered outcome improvements remain heterogeneous. Third, speculative or future-facing technologies-particularly AI-driven predictive or control systems and EIT-based perfusion applications-require further prospective validation, external testing, and implementation safeguards before routine adoption.

GAPS IN MONITORING AND EMERGING PHYSIOLOGIC TOOLS

A central barrier to individualized ventilation is that standard ventilator displays provide only a partial view of the relevant physiology. Airway pressure and flow reflect respiratory system behavior but cannot separate lung from chest wall mechanics, quantify inspiratory muscle pressure, or identify regional overdistension and collapse[6-10]. Consequently, clinicians often infer injury risk from indirect surrogates (tidal volume, plateau pressure, driving pressure, respiratory rate) without direct assessment of transpulmonary pressure, effort, or regional distribution[10-12].

Esophageal manometry can estimate pleural pressure and thereby approximate transpulmonary pressure, enabling clinicians to interpret airway pressures in the context of chest wall mechanics and to titrate PEEP with more explicit attention to lung distending pressure[8-10]. Although placement and interpretation require expertise and may be limited by artifacts, esophageal pressure monitoring is one of the few bedside approaches that directly addresses the lung-chest wall distinction.

Complementary tools focus on respiratory drive and effort. P0.1 provides a simple bedside estimate of central respiratory drive, but it can be influenced by ventilator algorithms, intrinsic (auto-) PEEP, respiratory mechanics, and neuromuscular weakness, so it should be interpreted alongside other effort measures and clinical context. At the bedside, selection of monitoring tools should be guided by the specific physiologic question being addressed. For example, EAdi is most useful for assessing neural drive and patient–ventilator interaction, esophageal pressure monitoring (Pes) is particularly valuable when distinguishing lung from chest wall mechanics or estimating transpulmonary pressure, and EIT may provide insight into regional ventilation distribution and recruitment. EAdi obtained via an esophageal electrode array, offers continuous measurement of neural activation of the diaphragm and can suggest under-assistance (high EAdi, high effort) or possible over-assistance (very low EAdi, risk of diaphragm disuse). However, low EAdi is not synonymous with over-assistance; if EAdi is low, first reassess sedation/analgesia and acid–base status, confirm catheter position/signal quality, and consider diaphragm dysfunction before reducing support[6,7,13-15]. However, availability of these monitoring modalities varies widely across institutions, and their use may be limited by equipment, expertise, and workflow considerations in many ICUs. Diaphragm ultrasound is also increasingly used to track diaphragm thickening and excursion, though standardized thresholds and outcome-driven protocols remain under development[33,54,55].

Regional monitoring is another key gap. EIT provides noninvasive bedside imaging of regional ventilation and, with specialized techniques, can be used to infer perfusion; perfusion applications are less standardized and not universally available, supporting assessment of recruitment, overdistension, pendelluft, and the distributional consequences of PEEP, positioning, or assisted breathing[11,12,56].

While EIT for regional ventilation assessment is increasingly used in specialized centers, perfusion-based applications remain investigational, are not yet standardized, and should not be considered routine decision-making tools. By transforming ventilation management from a global to a regional framework, EIT may help clinicians avoid strategies that improve oxygenation at the expense of localized overdistension or cyclic collapse.

Finally, composite metrics such as mechanical power integrate multiple ventilator variables into a single estimate of energy delivered per unit time. Mechanical power may capture risk that is not apparent from tidal volume or plateau pressure alone, but it is sensitive to ventilator mode, flow pattern, and patient effort and should be interpreted alongside measures of lung size, recruitability, and effort[11,57].

EMERGING NEURAL AND PROPORTIONAL MODES: NAVA AND PAV⁺

What problems do proportional modes solve that conventional ventilation does not?

Neurally adjusted ventilatory assist (NAVA) is a proportional mode in which inspiratory assist is coupled to the electrical activity of the diaphragm measured via an esophageal EAdi[13-15]. At the bedside, proportional modes such as NAVA and PAV⁺ may be particularly useful in patients with high respiratory drive, frequent patient–ventilator asynchrony, or difficulty achieving synchrony with conventional pressure- or volume-targeted modes. Appropriate patient selection is important, as these modes require an intact and reliable respiratory drive, absence of significant neuromuscular weakness, and the ability to generate consistent inspiratory effort. Triggering and cycling are based on neural timing rather than airway pressure/flow, and the magnitude of assist is scaled to EAdi via a clinician-set gain (“NAVA level”). Because neural activation precedes changes in airway pressure or flow, NAVA can reduce trigger delay and improve cycling synchrony in patients with intrinsic PEEP, weak inspiratory effort, or variable demand[31,58,59].

Multiple studies have demonstrated improvements in patient–ventilator interaction with NAVA (lower asynchrony index and fewer severe asynchrony events) compared with conventional pressure support ventilation[13,60]. However, evidence for patient-centered outcomes such as mortality or ICU length of stay remains uncertain and heterogeneous, likely reflecting variation in populations, co-interventions (sedation, weaning strategy), and the extent to which NAVA is used as a comprehensive lung- and diaphragm-protective strategy rather than as a single mode change[13,61]. Despite improvements in synchrony and patient-ventilator interaction, these physiologic benefits have not consistently translated into improved clinical outcomes, likely reflecting the multifactorial nature of critical illness. NAVA may facilitate lighter sedation in selected patients by improving comfort and synchrony, but sedation requirements are multifactorial and should not be attributed to ventilator mode alone[62].

Implementation considerations for NAVA include reliable EAdi acquisition, catheter positioning, recognition of signal artifact (e.g., electrical noise, low signal in profound weakness), and clear criteria for switching to backup modes when neural drive is absent or unstable. NAVA may be less effective when respiratory drive is profoundly suppressed, and it is not a substitute for addressing underlying causes of distress or acidosis that generate excessive neural output[14,15].

Proportional assist ventilation with load-adjustable gain factors (PAV⁺) is another proportional mode that delivers assistance proportional to estimated patient effort derived from measured flow and volume, using model-based estimates of resistance and elastance[15,63,64]. PAV⁺ is designed to improve synchrony, with some studies suggesting benefits in weaning efficiency compared with fixed pressure support, though outcome effects are not uniform across settings[64-66]. Because PAV⁺ relies on accurate estimates of mechanics, its performance can be affected by leaks, rapid changes in resistance/elastance, dynamic hyperinflation, and patient-ventilator circuit issues[63,64,67].

Overall, proportional modes are best viewed as enabling technologies for lung- and diaphragm-protective targets: They can improve coupling between patient demand and delivered support and may reduce unnecessary sedation or prolonged controlled ventilation. However, they do not eliminate the need for careful titration of assist, monitoring of tidal volume and effort, and management of underlying disease drivers of high respiratory drive[13-15,64].

AUTOMATION, CLOSED-LOOP VENTILATION, AND AI-ENABLED DECISION SUPPORT

What decisions can automation safely support, and where must clinician judgment remain central?

Automation in mechanical ventilation exists on a spectrum. At one end, rule-based closed-loop modes such as adaptive support ventilation and SmartCare/PS adjust support based on monitored variables (e.g., respiratory rate, tidal volume, end-tidal CO₂) to maintain predefined targets, often with a focus on standardizing support and streamlining weaning[68,69]. These systems can reduce the number of manual ventilator adjustments and may shorten weaning in selected populations, but effects on ICU length of stay, reintubation, and mortality have been inconsistent across trials and meta-analyses[69].

ML approaches differ in that they learn patterns from data and can be deployed for continuous monitoring, risk prediction, or, more speculatively, control policy generation[15-17,70,71]. A critical practical distinction is whether the tool is: (1) Descriptive (classifies events such as asynchrony); (2) Predictive (estimates a future risk such as extubation failure); or (3) Prescriptive/control-oriented (recommends or implements ventilator changes). Inputs to these systems may include ventilator waveform data, respiratory rate, tidal volume, airway pressure, and electronic health record variables such as laboratory values or vital signs. Outputs may include detection of asynchrony events, predicted risk of extubation failure, or recommendations for ventilator adjustments. Clinically, these outputs are most often integrated as decision-support prompts that guide ventilator titration, sedation adjustment, or escalation of monitoring rather than fully automated control actions.

The evidentiary and safety requirements increase substantially across this spectrum. At present, most AI applications in mechanical ventilation remain decision-support or retrospective analytic tools; autonomous control systems have not undergone sufficient prospective evaluation for routine clinical deployment. Decision-support systems provide recommendations or alerts that require clinician interpretation, whereas control systems directly adjust ventilator settings. For example, an asynchrony detection algorithm may alert clinicians to ineffective efforts (decision support), while a closed-loop ventilation mode may automatically adjust pressure support within predefined limits (control system).

Automated detection of patient-ventilator asynchrony and potentially injurious breathing patterns is among the most advanced ML applications in ventilation. Deep learning models trained on waveform datasets can identify ineffective efforts, double triggering, and cycling asynchronies with performance comparable to expert annotation in retrospective datasets[16,17]. If prospectively validated, such systems could support continuous monitoring, shorten time to recognition, and enable targeted ventilator optimization (trigger/cycle adjustments, assist level changes), potentially reducing unnecessary sedative escalation and limiting P-SILI risk. However, current studies vary substantially in labeling definitions, sampling frequency, patient populations, and reporting standards, which limits external validity and complicates benchmarking[72-74]. These differences in study design and implementation likely contribute to variability in reported outcomes.

Predictive analytics represent a second domain. Models combining physiologic data, laboratory values, and electronic health record features have been developed to estimate deterioration risk, ARDS trajectory, need for escalation (including extracorporeal membrane oxygenation), and likelihood of successful weaning/extubation[73,75,76]. Although many studies report high discrimination, bedside value depends on calibration (avoiding systematic over- or underestimation), interpretability, and integration into clinical workflows. Moreover, because outcomes such as extubation failure are influenced by clinician behavior, a model trained on observational data may learn local practice patterns rather than transportable physiology.

For AI systems that inform or automate ventilator changes, safety architecture is paramount. In addition to physiologic safety constraints, regulatory and medico-legal considerations are critical, including requirements for transparency in model inputs and outputs, validation across diverse populations, and clear assignment of clinical responsibility when automated recommendations or adjustments are made. Consensus statements emphasize predefined physiologic boundaries (e.g., limits on tidal volume, airway pressure, FiO₂, and respiratory rate), continuous anomaly detection, straightforward clinician override, and comprehensive audit trails to support quality improvement and accountability[77,78]. Systematic reviews highlight common risks in critical care AI: Retrospective single-center development, dataset shift, incomplete reporting, and limited external validation[72,74,79]. In the near term, AI is therefore most defensible as decision support and monitoring augmentation-helping clinicians quantify effort, detect asynchrony, and recognize deterioration- where must clinician judgment remain central[74,77-79].

CLINICAL APPLICATION: PRECISION VENTILATION IN ARDS AND SHOCK

Population-based lung-protective ventilation (approximately 6 mL/kg predicted body weight, limitation of plateau pressure, and avoidance of excessive driving pressure) improves outcomes in ARDS, as demonstrated in landmark randomized trials[6,47,48,50,80]. However, the syndrome is heterogeneous with respect to lung size, morphology, recruitability, chest wall mechanics, and cardiopulmonary reserve. Precision ventilation seeks to preserve core lung-protective principles while individualizing tidal volume, PEEP, and adjuncts (e.g., prone positioning) using bedside information on mechanics, recruitability, oxygenation response, and hemodynamic tolerance[48,50,51].

In this paradigm, monitoring of inspiratory effort becomes central. Excessive spontaneous effort may increase transpulmonary pressure and worsen regional overdistension or pendelluft, whereas complete suppression of effort can promote VIDD and prolong weaning[6,7]. Assisted modes-including proportional modes-may help target a moderate level of effort when paired with careful titration of assist, trigger/cycling settings, and sedation strategy. When excessive drive persists despite ventilator optimization and treatment of reversible contributors, short periods of deeper sedation or neuromuscular blockade may be appropriate to enable safe ventilation during the most unstable phases, particularly early severe ARDS with refractory hypoxemia or severe dyssynchrony[7,45,46].

In patients with shock or RV dysfunction, ventilatory decisions must be coordinated with hemodynamic assessment. Higher PEEP and mean airway pressure can impair venous return and increase RV afterload; therefore, PEEP titration should be individualized not only to oxygenation and recruitability but also to blood pressure, vasopressor requirements, and echocardiographic evidence of RV strain[49-51]. Prone positioning can improve oxygenation and redistribute stress/strain without necessarily increasing mean airway pressure and may therefore be particularly valuable when hemodynamic tolerance to higher PEEP is limited[48,50,51]. Similarly, fluid management in sepsis and ARDS requires balancing perfusion goals against the risk of worsening lung edema and impaired mechanics[34,53].

In practice, a stepwise bedside approach can be used to guide ventilator titration: (1) Establish lung-protective ventilation with appropriate tidal volume and plateau/driving pressure targets; (2) Assess oxygenation and recruitability to guide PEEP and adjuncts such as prone positioning; (3) Evaluate inspiratory effort using clinical assessment and available monitoring tools to avoid both excessive effort and over-assistance; and (4) Integrate hemodynamic response, including blood pressure, vasopressor requirements, and RV function, when adjusting ventilator settings. These domains should be interpreted together, as respiratory effort, lung mechanics, and hemodynamic status interact to determine the overall safety and effectiveness of a given ventilatory strategy. For example, in a patient with ARDS and concurrent shock, increasing PEEP to improve oxygenation may worsen hypotension and RV loading, prompting consideration of prone positioning or adjustment of ventilatory support to reduce injurious effort as an alternative strategy.

Taken together, this framework is intended to inform current bedside practice rather than represent a purely conceptual model. It supports a shift from reliance on isolated ventilator targets toward integrated assessment of respiratory effort, lung mechanics, regional ventilation, and hemodynamic response when titrating support. This approach emphasizes iterative reassessment of these interacting domains when adjusting ventilator settings, sedation, or adjunctive therapies, with the goal of maintaining lung- and diaphragm-protective ventilation within physiologic bounds. This process can be conceptualized as a cycle of assessment, adjustment, and reassessment during ongoing ventilator management.

IMPLEMENTATION CONSIDERATIONS

Common pitfalls and troubleshooting

Common troubleshooting considerations can be grouped into patient-related factors (e.g., respiratory drive abnormalities, neuromuscular weakness), ventilator-related factors (e.g., inappropriate settings, circuit leaks, trigger or cycling mismatch), and monitoring-related factors (e.g., signal quality, artifact, or device limitations).

EAdi/NAVA: If EAdi is low/absent or unstable, verify catheter position and signal quality (electrical noise/contact issues), reassess sedation depth and physiologic suppression of drive, and consider profound neuromuscular weakness. Ensure backup ventilation settings are explicit (e.g., apnea time/backup triggers) and document criteria for when to switch out of NAVA.

PAV⁺: If assist appears erratic or unexpected, evaluate for air leaks (cuff/circuit), rapid changes in resistance/elastance (secretions, bronchospasm), and auto-PEEP/dynamic hyperinflation; re-check mechanics estimates and consider reverting to a conventional assisted mode if mechanics are unstable.

Esophageal Pes: Interpret Pes in the context of chest wall mechanics and focus on trends/swings rather than overconfidence in single absolute values. Confirm placement/validity with standard checks (e.g., appropriate waveform behavior/occlusion-type validation) and recognize common artifact sources before acting on transpulmonary calculations.

EIT: Emphasize regional trends over single snapshots; troubleshoot belt position and electrode contact, and recognize motion/proning/contact artifacts that can mimic physiologic change. Avoid perfusion claims unless the specific methodology and validation support it.

Successful adoption of neurally informed modes and AI tools is determined as much by human factors and workflow as by technology. In practice, common barriers include clinician training and familiarity, equipment availability and cost, and time constraints in high-acuity ICU settings. For NAVA, best practice includes standardized catheter placement, routine EAdi signal-quality checks, and explicit criteria for initiation, titration, troubleshooting, and discontinuation[12,13]. Ventilator rounds can incorporate effort-focused targets (e.g., avoiding both very low and very high EAdi) alongside traditional oxygenation and pressure targets, reinforcing a lung- and diaphragm-protective framework.

For AI decision support, staged implementation (silent mode → clinician-facing decision support → limited automation) can mitigate alert fatigue and build clinician trust while maintaining accountability[77,78]. When decision support is presented, it should be contextualized (which variables drove the recommendation), bounded (what safety constraints are applied), and easy to override. Clear ownership, including who reviews AI output and when, prevents diffusion of responsibility and reduces the risk that alerts are ignored or misapplied.

From an infrastructure perspective, scalable deployment requires reliable acquisition of high-frequency ventilator data, interoperability with bedside monitors and the electronic health record, and governance frameworks that define oversight, privacy safeguards, cybersecurity, and monitoring for bias and model drift[78,79,81]. These requirements are prerequisites for safe translation of research prototypes into ICU practice and for meaningful prospective evaluation of clinical impact.

Limitations, barriers to adoption, and ethical considerations

Despite rapid technological progress, several barriers limit translation of advanced monitoring, proportional modes, and AI tools into routine ICU practice. First, signal acquisition can be invasive (esophageal pressure/EAdi) or resource-intensive (EIT), and performance may degrade with leaks, motion artifact, poor electrode contact, or severe neuromuscular weakness. Second, clinicians require training not only in device operation but also in physiologic interpretation; without clear protocols, additional data streams may increase cognitive load rather than improve decision-making[14,15,78].

AI tools introduce additional concerns. Models trained on retrospective data may embed local practice patterns and biases, and their performance can drift when patient populations, ventilator hardware, waveform sampling frequency, or clinical protocols change. In high-stakes applications, lack of transparency about model inputs and failure modes can impede safe use, while overly sensitive alerts can contribute to alarm fatigue and disregard of clinically important warnings[74,78,79]. Ethically, ventilation decision support must preserve patient-centered goals and clinician accountability. Consensus recommendations emphasize that clinicians and institutions remain responsible for decisions informed by AI, that models should be audited for bias and disparate performance, and that patients’ data privacy and security must be protected throughout data capture, storage, and model deployment[77,78]. Equity considerations are also important, as access to advanced monitoring tools and AI-enabled systems may vary across institutions, potentially widening disparities in care between resource-rich and resource-limited settings. In addition, formal cost-effectiveness data for many of these technologies remain limited, and their impact on resource utilization and long-term outcomes is not yet well defined.

Accordingly, implementation should be coupled to robust evaluation frameworks (prospective validation, calibration testing, monitoring for unintended consequences) and to transparent governance structures that define oversight and escalation pathways[78,79,81].

Future directions

Future ventilator platforms will likely integrate multimodal monitoring (airway pressure/flow, gas exchange, effort signals such as EAdi or esophageal pressure, and regional imaging such as EIT) with computational models that estimate patient-specific stress/strain and predict response to ventilator adjustments[4,82-85]. Computational physiology models and “virtual patient” frameworks may enable in silico testing of PEEP, tidal volume, or support changes before they are applied at the bedside, potentially reducing trial-and-error in unstable patients[83,85].

An important complementary direction is integration beyond the ventilator itself. Ventilatory settings are frequently adjusted in response to hemodynamic status, gas exchange trends, and imaging findings; yet these data streams are typically interpreted separately. Emerging platforms aim to fuse ventilator waveform data with continuous pulse oximetry, capnography, invasive arterial pressure, echocardiography-derived indices (when available), and laboratory trajectories. These systems may support real-time assessment of trade-offs between oxygenation, CO₂ clearance, lung stress, and perfusion. Such integrated monitoring approaches could be particularly valuable during transitions (initiation of spontaneous breathing trials, changes in PEEP, initiation of prone positioning, or liberation from extracorporeal membrane oxygenation), when both respiratory and cardiovascular instability may occur. Rigorous validation will be required to ensure that integrated dashboards improve clinical decisions rather than merely increasing data volume.

Data-driven control strategies such as reinforcement learning have shown proof-of-concept potential in retrospective or simulated environments, but clinical readiness depends on transparent objectives, robust external validation, and fail-safe design[85]. Even if control is partially automated, high-stakes decisions will continue to require clinician goal-setting (e.g., acceptable oxygenation/CO₂ targets, hemodynamic priorities), real-time assessment of patient comfort and trajectory, and accountability for complications. Thus, in the near term, the most plausible trajectory is incremental automation that supports, rather than replaces, clinician judgment in the delivery of life-supporting therapy. Longer-term developments may include more autonomous control strategies and fully integrated physiologic modeling platforms. However, these approaches remain experimental and require rigorous external validation, transparent objective functions, and fail-safe safeguards before clinical implementation. Key research priorities include prospective outcome-driven trials evaluating physiology-guided ventilation strategies and validation of integrated monitoring approaches across diverse patient populations and clinical settings (Figure 1).

Figure 1 Integrated physiology-guided ventilation framework. Schematic representation of a clinician-directed precision ventilation model. Central to the framework is integration of four physiologic domains: Respiratory effort monitoring (airway occlusion pressure, diaphragm electrical activity, esophageal pressure), lung stress and strain assessment (driving pressure, transpulmonary pressure, mechanical power), regional assessment of ventilation heterogeneity (electrical impedance tomography), and cardiopulmonary interactions (right ventricular preload and afterload, pleural pressure effects, and positive end-expiratory pressure-related hemodynamic consequences). These domains converge within a clinician-directed decision process rather than functioning as isolated targets. Bounded automation and artificial intelligence decision support, including waveform analysis and closed-loop control within predefined safety limits, serve as adjunctive tools. The overall objective is dynamic balancing of lung protection, diaphragm preservation, patient comfort, and hemodynamic stability in critically ill patients. AI: Artificial intelligence; PEEP: Positive end-expiratory pressure; P0.1: Airway occlusion pressure; EAdi: Diaphragm electrical activity; RV: Right ventricular.

CONCLUSION

Emerging technologies in mechanical ventilation are converging on a common goal: Safer, more individualized support that accounts for lung mechanics, respiratory drive, diaphragm function, and hemodynamics over time. At present, several elements of this approach are actionable at the bedside, including careful assessment of inspiratory effort, interpretation of airway pressures in the context of chest wall mechanics, recognition and correction of patient-ventilator asynchrony, and individualized titration of ventilatory support using available physiologic signals. These principles can be conceptualized as a physiology-guided framework that integrates respiratory effort monitoring, lung stress assessment, regional ventilation evaluation, and tailored ventilatory assist, supported by bounded decision-making that preserves clinician oversight (Figure 1). Future advances in physiologic monitoring and AI-enabled decision support may further refine this approach, but their role will depend on rigorous validation and thoughtful integration into clinical workflows. Realizing these benefits will require rigorous validation, careful implementation, and a shift from protocol-driven ventilation toward a dynamic, physiology-guided approach that maintains clinician oversight.