Validation of an automated patient-ventilator asynchrony detection algorithm

Background: Patient-ventilator asynchrony (PVA) is common in ventilated patients and associated with poor outcomes. However, it is underrecognized due to challenging identification. Automated algorithms aid, but are limited to few PVA types, lack validation, or operate offline [Rietveld et al, ICMx 2025, PMID:40119215]. Aim: To validate an AI-based algorithm (Deep Breath) that detects multiple PVAs in real-time (ineffective effort, reverse/double/delayed trigger and premature/delayed cycling). Methods: Ventilator and esophageal pressure (Pes) waveforms from 24 ventilated patients were collected. 5169 Breaths were labeled for PVA by 3 clinicians, out of a pool of 12, via dedicated software (Figure 1). Ground truth was set by majority voting or consensus. The same breaths were classified by the algorithm, using only ventilator waveforms (no Pes). We evaluated performance using recall, specificity and F1-score. Validation required ≥0.95 normal breath specificity and ≥0.8 recall per PVA type. Results: The algorithm accurately detects PVA, with a macro recall, specificity and F1-score of 0.83, 0.96 and 0.81, respectively. Minimum PVA recall (0.8, delayed trigger) and normal breath specificity (0.97) were above set validation criteria (Table 1). Conclusion: This novel algorithm is capable of real-time, accurate detection of multiple PVA types using only standard ventilator waveforms, offering a versatile solution for accurate PVA monitoring. erjor;12/suppl_17/44/F1 F1 F1