A machine learning approach to assess magnitude of asynchrony breathing. (April 2021)

• Record Type: Journal Article
• Authors: Loo, N.L.; Chiew, Y.S.; Tan, C.P.; Mat-Nor, M.B.; Ralib, A.M.
• Is Part Of: Biomedical signal processing and control. Volume 66(2021)
• Publication Date: 2021-04
• Subjects: Convolutional autoencoder (CAE) -- Real-time -- Asynchronous breathing -- Patient-ventilator interaction; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102505
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Magnitude of asynchrony breathing assessment during MV is underrecognized. A machine learning approach is presented to reconstruct asynchrony breathing. Model allows quantification of magnitude of asynchrony breathing. Magnitude analysis on clinical data is significantly lower than AI computation. Potentially provide a better indication of the quality of MV delivery. Abstract: Background: Conventional patient-ventilator interaction (PVI) assessment involves manual asynchronous index (AI) computation and incapable to provide in-depth information of the severity of asynchrony breathing (AB) during mechanical ventilation (MV). In this study, a novel convolutional autoencoder model ( ABReCA ) is developed to quantify the magnitude of AB as indicator of PVI. Method: ABReCA was trained with 400.000 unique AB to recognise its corresponding normal breathing (NB) cycle. The model then quantifies the severity of AB through comparison between identified NB waveform and AB. The magnitude of asynchrony (Masyn) is defined as the difference of a NB cycle affected by asynchronous patient's effort. The performance of ABReCA was evaluated using K-folds analysis and used to measure the severity of AB in 10 mechanical ventilated respiratory failure patients. Results: K-fold analysis showed that ABReCA achieved high performance with only median 0.008 [Interquartile range (IQR): 0.007−0.010] validation error. The model was able to recognise AB and its corresponding NB cycle. For the actual MV patient analysis, a typical AI counter shows a median of 32.7 % [IQR: 32.1–34.4] per patient. However, in our magnitude analysis, these patients experienced Masyn with mean of 3.8 % [IQR: 1.7 %–4.6 %]. The severity result is significantly lower compared to counting numbers alone as some AB are negligible while others have more impact towards the overall MV delivery. Conclusion: A novel ABReCA is developed and capable of quantifying the severity of AB during MV. This model can potentially provide a better indication of the severity of AB and better reflection of the quality of PVI.