Medical percussion is a free, low-risk procedure used by physicians during physical examination of patients. Although it is very useful procedure, a downside to manual percussion is that its results are subjective, with typically low inter-observer agreement. Not much work has been done, however, to create automated and reliable percussion devices or percussograph. This paper reports the development of a mobile percussograph. A spring-loaded solenoid was used as the plessor generating mechanical impact for application to a subject’s skin. Generated signals were amplified and digitized at a rate of 22.1 kHz. Thereafter, Frequency B-Spline (FBSP) base wavelet transform at 512 scales was used for feature extraction. Spectrographs generated from the wavelet coefficients were used for training a MobileNet network with 17 inverted layers for a 3-way classification. Training employed a cross entropy loss function and the Adam optimization algorithm. Learning rate was 0.001, and first and second moment decay rates were 0.9 and 0.999 respectively. Subject-specific test accuracies of 92.9 %, 93.7 %, and 96.4 % were obtained for three subjects, while the cross-subject classification accuracy was 95.0 %. As this is the first reported general purpose mobile percussograph reported in the literature, these results are state-of-the-art. This study has established the viability of implementing mobile percussography in a standard, repeatable and accurate manner, which can lead to faster and more reliable medical percussion globally.Keywords— MobileNet, Percussion, Percussograph, Percussography, Wavelets
A Mobile Percussograph for Medical Examination of the Torso