Effects of Data Structure in Convolutional Neural Network for Detection of Asynchronous Breathing in Mechanical Ventilation Treatment

Abstract

Asynchronous breathing (AB) in mechanical ventilation (MV) patients is heterogenous, patient-specific, and is associated with adverse patient outcomes. Various machine learning models have been developed for AB detection, however studies regarding the data structures used for model training are scarce. This study investigates the effects of different training data structures and sizes of Convolutional Neural Networks (CNN) to detect AB. Four CNN models were developed using different amounts of data and data structures: one-dimension, line, area, and array. Training datasets consisting of 300, 1,000, 5,000 and 10,000 airway pressure waveforms from MV patients were used for model development. Model sensitivity and specificity were evaluated using an independent set of 3000 waveforms in a 100-iteration Monte-Carlo analysis. The best-performing CNN model was used to determine Asynchrony Index (AI) values in a clinical patient cohort. Monte-Carlo analysis showed that models trained with datasets of 10,000 breathing cycles delivered $\gt99$% sensitivity and specificity. Relatively lower sensitivity and specificity of $\lt78.8$% and $\lt96.7$% respectively were obtained when trained with data quantities of 5000 breaths or less. A CNN trained with 1Dimensional data structure yielded 99.9% sensitivity and 99.6% specificity. It achieved 88.5% average accuracy when validated with an independent clinical data set of 544,319 breaths. Asynchrony breathing detection is ubiquitous and 1-Dimensional data structures provide a resource efficient method for the development of an accurate CNN model.