| Popis: |
Cough is a common symptom of respiratory diseases and the type of cough, in particular, productive (wet) or non-productive (dry) cough, is an important indicator of the condition of the respiratory system. It is useful in differential diagnosis and in understanding disease progression. However, determining the cough type in clinical practice can be subjective and sometimes unfeasible. This work, therefore, aims to develop an objective assessment method of the cough type. The proposed approach emulates the sound recognition process of humans. In particular, it uses the human auditory model to reveal the frequency characteristics of the cough sound signals and convolutional neural networks for decision-making. It is validated on a dataset of smartphone recordings of 396 cough samples from 88 subjects annotated as wet or dry by up to four expert pulmonologists. The cough signals are automatically segmented and time-frequency image data augmentation is performed during training using the synthetic minority oversampling technique to prevent model overfitting. A sensitivity of 93.13% and specificity of 91.42% (AUC= 0.9700) is achieved in segmentation of cough and non-cough sounds and a sensitivity of 100% and specificity of 82.50% (AUC= 0.9234) is achieved in detecting subjects with wet and dry cough. The proposed fully automated system in detecting subjects with wet and dry cough demonstrates strong classification performance. It has the potential to provide objective assessment of cough type using smartphone technology, such as in virtual healthcare which has seen an increased uptake during the ongoing pandemic. |
