AI Algorithm Continuously Monitors Inspiratory Muscle Effort, Patient-Ventilator Synchrony

Advancing Patient-Ventilator Synchrony Monitoring in Critical Care

Patient-ventilator asynchrony (PVA) is a frequent and serious complication during mechanical ventilation, linked to prolonged ventilation, increased morbidity, and higher mortality in critically ill patients [1, 2, 3]. Despite its clinical importance, reliably detecting PVA at the bedside is difficult. Current practice often depends on intermittent visual inspection of ventilator waveforms, a method with low sensitivity even for experienced clinicians [4, 5, 6]. The dynamic and complex nature of patient-ventilator interaction means that many asynchronous events are missed, potentially leading to underdiagnosis and suboptimal ventilator management [7, 5]. In this context, computational tools using artificial intelligence (AI) are being investigated to improve diagnostic accuracy in critical care, particularly for managing mechanical ventilation and acute respiratory distress syndrome (ARDS) [8, 9, 10, 11]. A recent study evaluated an AI-based algorithm designed to provide continuous, objective monitoring of a patient's respiratory effort and synchrony with the ventilator [5, 2].

Addressing Limitations in Ventilatory Monitoring

Optimizing mechanical ventilation requires an accurate assessment of inspiratory muscle pressure (Pmus), which reflects the patient's work of breathing. However, current methods are constrained; they are either invasive or provide only intermittent data. The gold standard, esophageal manometry, offers precise, continuous measurements but requires placing a catheter, which involves patient discomfort and procedural risks, limiting its routine use. In contrast, non-invasive methods typically depend on temporary occlusion maneuvers. These maneuvers interrupt normal ventilation to provide a snapshot of patient effort, but they are time-consuming and cannot offer the continuous, real-time evaluation needed for dynamic patient management. To address these gaps, researchers developed and tested an artificial intelligence algorithm engineered to estimate the amplitude and timing of Pmus noninvasively and in real-time. The clinical objective is to furnish a continuous stream of data on patient effort, driving pressure, and synchrony, giving clinicians a more comprehensive view of patient-ventilator interaction without invasive lines or disruptive procedures.

Study Design and Patient Cohort

To validate the algorithm, the investigators conducted a prospective diagnostic accuracy study, a design that allows for direct comparison of a new test against an established gold standard in a clinical setting. The research was performed in two intensive care units (ICUs) at the University of São Paulo, Brazil, ensuring the findings are relevant to contemporary critical care practice. The study enrolled adult patients receiving pressure support ventilation, a common mode of partial ventilatory support where the patient triggers each breath. The final analysis included data from 48 participants, from whom a total of 4918 respiratory cycles were recorded and evaluated. This large dataset of individual breaths provided a robust foundation for assessing the AI algorithm's performance across a wide variety of clinical conditions and respiratory patterns.

Evaluating AI Performance Against Established Methods

The algorithm's accuracy was rigorously tested by comparing its real-time Pmus estimates (Pmus,AI) against two benchmarks. The primary comparison was with the invasive gold standard, esophageal manometry (Pmus,es). The study captured a broad spectrum of patient effort, with Pmus,es values ranging from 1.0 to 28.4 cm H2O, confirming the algorithm was tested under diverse physiological conditions. The findings showed a mean bias of 0.9 cm H2O between Pmus,AI and Pmus,es, indicating a small systematic overestimation by the algorithm. The 95% limits of agreement were -5.1 to 6.9 cm H2O, defining the expected range of error for most individual measurements. Critically, the algorithm was effective at identifying clinically important extremes of respiratory effort. Its ability to detect both excessively high and low Pmus,es and dynamic driving pressure, which are risk factors for lung and diaphragm injury, was strong, with an area under the receiver operating characteristic curve greater than 0.8. The study also found the accuracy of the continuous Pmus,AI was comparable to that of intermittent occlusion-based techniques, such as the pressure muscle index and occlusion pressure (Pocc).

Automated Dyssynchrony Detection

A key function of the algorithm is its potential to automate the detection of specific patient-ventilator dyssynchronies, which are often missed by visual inspection. The researchers compared the AI's automated classifications against a consensus of expert clinicians who reviewed the raw data. For identifying common and detrimental dyssynchronies like ineffective effort, autotriggering, or reverse triggering, the algorithm demonstrated a sensitivity of 86.5% and a specificity of 77.4%. The high sensitivity suggests the tool is reliable for flagging these events for clinical review, while the moderate specificity indicates that some false positives may occur, requiring clinician confirmation at the bedside. In their conclusions, the authors note the algorithm performed well in detecting both high and low Pmus, which is essential for identifying patients at risk of diaphragm injury from either over-assistance or excessive effort. This noninvasive method provides continuous monitoring of both respiratory effort and specific dyssynchronies, with performance comparable to established but intermittent and disruptive occlusion maneuvers.

References

1. Jiang G, Ma J, Hao X, et al. Application progress of machine learning in patient-ventilator asynchrony during mechanical ventilation: a systematic review. Critical Care. 2025. doi:10.1186/s13054-025-05523-3

2. Docci M, Rodrigues A, Dubó S, Ko M, Brochard L. Does patient-ventilator asynchrony really matter?. Current Opinion in Critical Care. 2024. doi:10.1097/mcc.0000000000001225

3. Docci M, Rodrigues A, Dubó S, Ko M, Brochard L. Does patient-ventilator asynchrony really matter?. Current Opinion in Critical Care. 2024. doi:10.1097/MCC.0000000000001225

4. Bandeira M, Almeida A, Silva LGMD, et al. Accuracy of Algorithms and Visual Inspection for Detection of Trigger Asynchrony in Critical Patients : A Systematic Review. Critical Care Research and Practice. 2021. doi:10.1155/2021/6942497

5. Sarlabous L, Suñol F, Magrans R, et al. Automated Detection of Ventilator Asynchronies: A Clinical and Technological Perspective. Respiratory Care. 2026. doi:10.1177/19433654251412329

6. Silva DO, Souza PND, Sousa MLDA, et al. Impact on the ability of healthcare professionals to correctly identify patient-ventilator asynchronies of the simultaneous visualization of estimated muscle pressure curves on the ventilator display: a randomized study (Pmus study). Critical Care. 2023. doi:10.1186/s13054-023-04414-9

7. Haro CD, Santos-Pulpón V, Telías I, et al. Flow starvation during square-flow assisted ventilation detected by supervised deep learning techniques. Critical Care. 2024. doi:10.1186/s13054-024-04845-y

8. Lombardi R, Chaturvedi M, Jozwiak M, Garg M. Machine learning in ARDS: an intensivist’s guide to artificial intelligence applications. Critical Care. 2026. doi:10.1186/s13054-026-05928-8

9. Pearce AK, Nemati S, Goligher EC, et al. Can we predict the future of respiratory failure prediction?. Critical Care. 2025. doi:10.1186/s13054-025-05484-7

10. Hadweh P, Niset A, Salvagno M, et al. Machine Learning and Artificial Intelligence in Intensive Care Medicine: Critical Recalibrations from Rule-Based Systems to Frontier Models. Journal of Clinical Medicine. 2025. doi:10.3390/jcm14124026

11. Karthika M, Sreedharan JK, Shevade M, Mathew CS, Ray S. Artificial intelligence in respiratory care. Frontiers in Digital Health. 2024. doi:10.3389/fdgth.2024.1502434