Phenotyping the physiological response to NAVA using non-linear modeling to identify patients at risk of patient-self inflicted lung injury

Background. Acute Respiratory Failure (ARF) often requires transition from controlled to assisted mechanical ventilation as patients stabilize. While this strategy may reduce ventilator-induced diaphragmatic dysfunction (VIDD) and ventilator-induced lung injury (VILI), it can increase the risk of patient self-inflicted lung injury (P-SILI), caused by excessive spontaneous inspiratory effort leading to high transpulmonary pressures and large tidal volumes (VT). Neurally Adjusted Ventilatory Assist (NAVA) delivers support proportionally to diaphragmatic electrical activity (EAdi), potentially improving patient–ventilator interaction. However, patient responses to NAVA are highly variable, and it remains unclear whether specific response patterns identify individuals at greater risk of lung injury. Aim of the study. The primary objective of the study is to characterize physiological heterogeneity during incremental NAVA titration using non-linear modeling and identify VT response patterns associated with respiratory drive, effort, and P-SILI risk. A secondary objective was to assess the association between these phenotypes and clinical outcomes. Materials and methods. In this prospective observational study, 47 adult patients at risk of difficult weaning were enrolled in two ICUs of Padova University Hospital. Following baseline assessment in Pressure Support Ventilation (PSV), NAVA was increased from 0.5 to 3.0 cmH₂O/µV in 0.5-step increments, each lasting 20 minutes. Respiratory variables included tidal volume normalized to ideal body weight (VT/IBW), respiratory rate, EAdi peak, airway occlusion pressure (ΔPocc), static driving pressure (DP), dynamic transpulmonary driving pressure (ΔPL,dyn), and neuromuscular efficiency (NME). Individual responses were modeled using cubic B-splines selected by the lowest Akaike Information Criterion (AIC), and patients were clustered according to VT trajectories through finite mixture models. Results. Forty-seven patients were analyzed, and cubic B-splines provided the best fit in 64% of cases. Three physiological phenotypes emerged: safe response (21%), with VT/IBW remaining within protective limits and reaching 6.10 (5.34–7.26) mL/kg; dangerous response (45%), characterized by a marked VT/IBW increase up to 10.35 (9.37–11.45) mL/kg and elevated lung stress indices, with DP 14.0 (8.0–16.0) cmH₂O and ΔPL,dyn 25.6 (16.4–36.7) cmH₂O; and intermediate response (34%), reaching 7.77 (6.95–8.69) mL/kg. Baseline clinical and arterial blood gas variables did not differentiate clusters, whereas VT behavior during titration accurately predicted phenotype membership (AUC > 0.93). Safe responders had shorter invasive mechanical ventilation duration (6 vs. 12 days; p=0.015) and ICU stay (8 vs. 15 days; p=0.029) than dangerous responders, while the intermediate phenotype showed the poorest overall prognosis. Conclusions. The physiological response to NAVA is highly heterogeneous in patients at risk of difficult weaning. This study demonstrates that clustering VT trajectories unmasks distinct physiological phenotypes with varying P-SILI risk profiles that are otherwise invisible to standard clinical monitoring. If validated in larger cohorts, this phenotyping approach could provide a reproducible bedside strategy for individualizing assisted ventilation and improving risk stratification. Identifying these patterns early may facilitate a transition toward precision medicine in the ICU, helping clinicians avoid unrecognized lung injury during spontaneous breathing.