Artificial Intelligence Algorithm to Monitor Inspiratory Muscle Effort and Patient-Ventilator Dyssynchrony During Mechanical Ventilation

OBJECTIVE: Current methods for estimating inspiratory muscle pressure ( Pmus ) during mechanical ventilation are either invasive or dependent on occlusion maneuvers. A noninvasive artificial intelligence (AI) algorithm estimating in real-time the amplitude and timing of Pmus , enabling continuous monitoring of patient effort, driving pressure, and synchrony with the ventilator was designed, and its performance was evaluated against the gold standard obtained with esophageal manometry ( Pmus,es ). DESIGN: A prospective diagnostic accuracy study. SETTING: Two ICUs from the University of São Paulo, Brazil. PATIENTS: Adult patients under pressure support ventilation. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Pmus estimated using AI ( Pmus,AI ) was compared with Pmus,es and to values derived from occlusion maneuvers, the pressure muscle index and the occlusion pressure ( Pocc ). Automatic detection of dyssynchronies based on Pmus,AI was compared with experts' classification. A total of 48 participants with 4918 cycles were analyzed. Pmus,es varied from 1.0 to 28.4 cm H 2 O. Pmus,AI showed a bias of 0.9 cm H 2 O, 95% limits of agreement -5.1, 6.9 cm H 2 O and detected extreme values of both Pmus,es and dynamic driving pressure with area under the receiver operating characteristic curve greater than 0.8. Pmus,AI accuracy was comparable to occlusion-based techniques. Sensitivity and specificity to detect ineffective effort, autotriggering or reverse triggering were 86.5% and 77.4%, respectively. CONCLUSIONS: AI presented good performance in detecting high and low Pmus , and allowed the automatic detection of specific types of dyssynchronies. This novel noninvasive method was comparable to intermittent techniques requiring occlusion maneuvers.