Utilizing Machine Learning and Esophageal Pressure to Detect Ventilator Dyssynchrony and Delineate the Potential for Ventilator Induced Lung Injury

Rationale: It is recognized that ventilator dyssynchrony (VD) may propagate ventilator induced lung injury (VILI). Yet some VD cannot be detected without advanced monitoring like measuring esophageal pressure (Pes) and it is unknown which types of VD propagate VILI. We describe the automated detection of VD using machine learning (ML) in patients with esophageal manometry to quantify the frequency and association between VD, tidal volume (VT) and transpulmonary driving pressure (ΔPdyn.tp). Methods: We enrolled 42 patients with ARDS or ARDS risk factors, including COVID-19. XGBoost, a ML algorithm, was trained to identify 7 types of breaths using a one-vs-all strategy, from a training set of 3500 random breaths. We compared the models' sensitivity and specificity with and without features derived from Pes. Finally, the association between each VD type and VT or ΔPdyn.tp was calculated using separate linear mixed-effect models. Temporally related breaths were nested by patient and modeled as random effects, accounting for repeat measures and changing pulmonary mechanics in each patient. Breaths without an adequate Pes signal were excluded from analysis Results: Patients were 37.5% female, 52±15 years old, had an initial P:F ratio of 140±64, and 24.2% of the 480, 976 breaths were dyssynchronous. Normal passive (Nlp), normal spontaneous (Nls), late reverse triggered (RTl), reverse triggered double triggered (DTr), mild flow limited (FLm), severe flow limited (FLs), and early ventilator terminated (EVT) account for 47.0%, 28.7%, 4.8%, 3.7%, 8.9%, 4.2%, and 2.5% of all breaths, respectively. ML training, VT and ΔPdyn.tp results are show in the table (∗p<0.001). Conclusion: ML algorithms can be trained using Pes to identify types of VD that traditionally need Pes measurements, although without Pes sensitivity may decrease. VD is frequent and DTr is associated with an increase in VT, while FLm and FLs are associated with an increased in ΔPdyn.tp. These data suggest that double triggered breaths and flow limited breaths have the potential to propagate VILI, while other types of VD may not be as deleterious. (Table Presented).