Generative Adversarial Network in Reconstructing Asynchronous Breathing Cycle

Asynchronous breathing (AB) during mechanical ventilation (MV) can have adverse effect towards a patient’s recovery. Especially, the presence of AB will disrupt MV breathing profile; thus, misidentifying patient-specific condition. This paper demonstrates the ability of generative adversarial network (GAN) to reconstruct asynchronous breaths to a normal breath profile. The reconstructed clean airway pressure can provide better identification of patient’s condition. A total of 120,000 asynchronous and normal breaths GAN training data set were simulated from a Gaussian effort model. The breaths consist of elastance from 15 to 35 cmH2O/L and resistance from 10 to 20 cmH2Os/L. Three GAN configurations were investigated in this study. The first GAN configuration trained with 120,000 breaths yielded error of median 6.0 cmH2O/L [interquartile range (IQR): 3.71-11.56]. The second configuration comprised of five GAN models improved with median error of 2.48 cmH2O/L [IQR: 1.19-4.69] with each model trained in five different elastance and resistance values. The third configuration had 15 GAN models with each model trained with one set of elastance and resistance. The median error was 0.70 cmH2O/L [IQR: 0.22-4.29] for the third configuration. The results indicate that by dissipating the classification task, the performance of GAN reconstructing AB can be improved. Realizing GAN in real-time to reconstruct AB to a normal breath can potentially improve patient’s condition diagnosis.