A 1-D Deformable Convolutional Neural Network for the Quantitative Analysis of Capnographic Sensor

Abstract

Capnography, the measurement of the concentration of carbon dioxide in exhaled air has the potential to provide insight into the diagnosis of many cardiopulmonary diseases. Our objective is to develop an effective deep learning algorithm to analyze capnograph and thereby diagnose Chronic Obstructive Pulmonary Disease (COPD) and Congestive Heart Failure (CHF). In this paper, we propose a novel 1D Deformable Convolutional Neural Network for the quantitative analysis of the capnograph. A learnable and deformable convolution layer and pooling layer are incorporated in 1D Convolutional Neural Network (CNN) to form a 1D Deformable CNN. The feature extraction and classification which are the main two steps in quantitative analysis of capnograph are fused into a single learning system in the proposed architecture. The learnable parameters are extracted by backpropagating the error. The proposed system classifies the capnographic signals extracted using a speed of sound-based sensor and the signal from a reference database into three classes namely, Healthy, COPD and CHF. The experimental results were clinically validated. Performance analysis yields an average classification accuracy of 92.9% and 92.16% for the signal from the reference database and the signal acquired using the sensor respectively. The performance depicts the feasibility of the proposed 1D Deformable CNN to replace existing methods for diagnosis using capnograph. The analysis shows that better learning is achieved in 1D Deformable CNN compared to state-of-the-art techniques. Moreover, no prior knowledge is required to use the system because the proposed method performs feature extraction and classification automatically.