Automatic quality assessment of capacitively-coupled bioimpedance signals for respiratory activity monitoring

The objective of this work is to design an algorithm capable of classifying segments of capacitively-coupled bioimpedance (ccBioZ) for respiratory activity monitoring based on their signal quality. Such algorithm is an important building block as a pre-processing step to increase the confidence of extracted information such as respiration rate (RR). Long over-night ccBioZ recordings acquired from 12 subjects, are used for training and testing the proposed algorithm. To create a ground-truth labelling, the annotation is done manually by an experienced biomedical engineer. A total number of 52 features are extracted to capture information related to the quality of the ccBioZ segments. Five subsets of features are selected based on five different feature selection methods and tested against a full set of features to find the best trade-off between the performance of the classifier and the number of features. For classification, 19 classifiers are trained, cross-validated, and tested in three different datasets, acquired from 12 subjects: DS1 (training and validation data from 11 patients with suspected sleep apnea), DS2 (containing apneic epochs acquired from the same 11 patients), and DS3 (testing data acquired from one healthy subject). The balanced accuracy is used along with other statistical evaluation metrics. For each test set, the best results of the quantitative evaluation came as following; DS1: Accuracy = 0.91, Sensitivity = 0.90, Specificity = 0.95, and Balanced Accuracy = 0.91. DS2: Accuracy = 0.87, Sensitivity = 0.88, Specificity = 0.87, and Balanced Accuracy = 0.88. DS3: Accuracy = 0.91, Sensitivity = 0.98, Specificity = 0.91, and Balanced Accuracy = 0.94. The results of the testing phases prove the reliability and robustness of the presented approach.

Automated summary

The objective of this study was to design an algorithm for classifying ccBioZ segments based on signal quality to enhance confidence in extracted respiratory activity monitoring information, such as respiration rate. This was achieved through extracting and selecting features to find the best balance between classifier performance and the number of features used. Testing the algorithm on three datasets from 12 subjects showed high accuracy, sensitivity, specificity, and balanced accuracy, indicating the reliability and robustness of the approach.