Artificial Intelligence for detecting asynchrony in non-invasive ventilation: a concordance study.

Introduction: Artificial intelligence (AI) is increasingly applied in respiratory medicine. Its role in detecting patient–ventilator asynchronies during non-invasive ventilation (NIV) remains unclear. Asynchrony is associated with worse outcomes and NIV failure (Longini F et al, Front Med; Ghosh D et al ERJ Open Research 2017). Current identification relies on expert visual inspection, which is subjective and time-consuming. Aim. To evaluate the agreement between AI platforms and expert pulmonologists in identifying patient–ventilator asynchronies during NIV. Materials and Methods. NIV waveforms were independently analysed by two experienced pulmonologists using SOMNONIV criteria (Gonzalez-Bermejo J, et al. Thorax 2019), retaining only traces with full inter-observer agreement. Two AI platforms (Perplexity® and ChatGPT-5®) were pretrained with 10 representative waveforms and then applied to 80 validated traces using a standardized prompt. Results. AI platforms showed poor agreement with pulmonologists. ChatGPT-5® slightly outperformed Perplexity®, though performance varied by software (table 1). Conclusions. AI currently shows poor agreement with pulmonologists in detecting asynchronies during NIV. ChatGPT5® performed slightly better, but results remain inconsistent and not suitable for clinical use. erjor;12/suppl_17/191/TB1 T1 TB1 Table 1. Concordance in asynchrony detection. Pulmonologist vs AI platform Kappa (95%CI) Global agreement Perplexity® -ResScan® -Orchestrator® -Breas® 0.11 (-0.03 – 0.25) 0.21 (-0.05 – 0.48) 0.02 (-0.13 – 0.18) -0.06 (-0.41 – 0.28) 47.5% 60.0% 37.5% 38.5% ChatGPT5® -ResScan® -Orchestrator® -Breas® 0.13 (-0.07 – 0.32) 0.30 (0.10 – 0.50) 0.13 (-0.23 – 0.48) -0.31 (-0.81 – 0.18) 56.2% 60.0% 62.5% 30.8%