A medical percussion instrument using a wavelet-based method for archivable output and automatic classification

There is no standard instrument for carrying out medical percussion even though the procedure has been in continuous use since 1761. This study developed one such instrument. It generates medical percussion sounds in a reproducible manner and accurately classifies them into one of three classes. Percussion signals were generated using a push-pull solenoid plessor applying mechanical impulses through a polyvinyl chloride plessimeter. Signals were acquired using a National Instruments USB 6251 data acquisition card at a rate of 8.192 kHz through an air-coupled omnidirectional electret microphone located 60 mm from the impact site. Signal acquisition, processing, and classification were controlled by an NVIDIA Jetson TX2 computational device. A complex Morlet wavelet was selected as the base wavelet for the wavelet decomposition using the maximum wavelet energy method. It was also used to generate a scalogram suitable for manual or automatic classification. Automatic classification was achieved using a MobileNetv2 convolutional neural network with 17 inverted residual layers on the basis of 224 × 224 x 1 images generated by downsampling each scalogram. Testing was carried out using five human subjects with impulses applied at three thoracic sites each to elicit dull, resonant, and tympanic signals respectively. Classifier training utilized the Adam algorithm with a learning rate of 0.001, and first and second moments of 0.9 and 0.999 respectively for 100 epochs, with early stopping. Mean subject-specific validation and test accuracies of 95.9±1.6% and 93.8±2.3% respectively were obtained, along with cross-subject validation and test accuracies of 94.9% and 94.0% respectively. These results compare very favorably with previously-reported systems for automatic generation and classification of percussion sounds. •A system has been developed for accurate, reproducible medical percussography.•A complex Morlet base wavelet leads to an output that can be interpreted manually or automatically.•A classifier based on standard MobileNetv2 architecture was used for automatic classification.•Mean subject-specific validation accuracy of 95.9±1.6% and subject-specific test accuracy of 93.8±2.3% were achieved.•These are higher than classification accuracies reported by previous studies.