Automatic Recognition and Classification of Patient–Ventilator Asynchrony Using Deep Learning with Generative Adversarial Networks

Purpose: Patient–ventilator asynchrony (PVA) is frequent in mechanically ventilated patients and is associated with poor outcomes. Automated recognition has the potential to improve clinical care, but existing methods are limited by data scarcity and class imbalance. We aimed to develop and validate a deep learning model enhanced by synthetic data generation to improve the accuracy of PVA detection and classification. Methods: We developed a two-step framework: (1) a self-attention CNN–LSTM generative adversarial network (SACL-GAN) to generate realistic respiratory cycles, and (2) a hybrid multi-head attention CNN–BiLSTM classifier (MHACBL) model trained on both real and synthetic data. 4,907 respiratory cycles from 32 ICU patients were annotated into normal breathing (NB) and four asynchrony categories: ineffective effort (IE), early cycling (EC), double triggering (DT), and airway oscillation (AO). Model performance was assessed using accuracy, precision, recall, and F1-score and other metrics. Results: SACL-GAN successfully generated high-quality synthetic data with low mean squared error (MSE: 0.58-2.84 for pressure, 1.68-8.07 for flow) and dynamic time warping (DTW: 3.03-6.47 for pressure, 9.61-12.24 for flow). Without augmentation, the MHA-CBL classifier achieved high accuracy for NB and IE, but lower recall for minority classes (EC and AO). GAN augmentation improved overall accuracy (from 94.6% to 97.8%) and recall for minority classes (EC: 76.9% → 93.2%; AO: 74.6% → 91.8%). In binary classification (asynchrony vs NB), sensitivity improved from 94.4% to 98.0%, reducing missed detections. Conclusion: Combining GAN-based augmentation with hybrid deep learning improved recognition of PVA, particularly for underrepresented categories. This strategy addresses dataset imbalance and enhances the clinical applicability of automated monitoring in ICU settings.