A physical approach to the automated classification of clinical percussion sounds

Telemedical percussion: objectifying a fundamental clinical examination technique for telemedicine

Purpose

While demand for telemedicine is increasing, patients are currently restricted to tele-consultation for the most part. Fundamental diagnostics like the percussion still require the in person expertize of a physician. To meet today’s challenges, a transformation of the manual percussion into a standardized, digital version, ready for telemedical execution is required.

Methods

In conjunction with a comprehensive telemedical diagnostic system, in which patients can get examined by a remote-physician, a series of three robotic end-effectors for mechanical percussion were developed. Comprising a motor, a magnetic and a pneumatic-based version, the devices strike a pleximeter to perform the percussion. Emitted sounds were captured using a microphone-equipped stethoscope. The 84 recordings were further integrated into a survey in order to classify lung and non-lung samples.

Results

The study with 21 participants comprised physicians, medical students and non-medical-related raters in equal parts. With 71.4% correctly classified samples, the ventral motorized device prevailed. While the result is significantly better compared to a manual or pneumatic percussion in this very setup, it only has a small edge over the magnetic devices. In addition, for all ventral versions non-lung regions were rather correctly identified than lung regions.

Conclusion

The overall setup proves the feasibility of a telemedical percussion. Despite the fact, that produced sounds differ compared to today’s manual technique, the study shows that a standardized mechanical percussion has the potential to improve the gold standard’s accuracy. While further extensive medical evaluation is yet to come, the system paves the way for future uncompromised remote examinations.

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.

Time-Frequency Analysis of Clinical Percussion Signals Using Matrix Pencil Method

This paper discusses time-frequency analysis of clinical percussion signals produced by tapping over human chest or abdomen with a neurological hammer and recorded with an air microphone. The analysis of short, highly damped percussion signals using conventional time-frequency distributions (TFDs) meets certain difficulties, such as poor time-frequency localization, cross terms, and masking of the lower energy features by the higher energy ones. The above shortcomings lead to inaccurate and ambiguous representation of the signal behavior in the time-frequency plane. This work describes an attempt to construct a TF representation specifically tailored to clinical percussion signals to achieve better resolution of individual components corresponding to physical oscillation modes. Matrix Pencil Method (MPM) is used to decompose the signal into a set of exponentially damped sinusoids, which are then plotted in the time-frequency plane. Such representation provides better visualization of the signal structure than the commonly used frequency-amplitude plots and facilitates tracking subtle changes in the signal for diagnostic purposes. The performance of our approach has been verified on both ideal and real percussion signals. The MPM-based time-frequency analysis appears to be a better choice for clinical percussion signals than conventional TFDs, while its ability to visualize damping has immediate practical applications.

Advantages and limitations of using matrix pencil method for the modal analysis of medical percussion signals

Although clinical percussion remains one of the most widespread traditional noninvasive methods for diagnosing pulmonary disease, the available analysis of physical characteristics of the percussion sound using modern signal processing techniques is still quite limited. The majority of existing literature on the subject reports either time-domain or spectral analysis methods. However, Fourier analysis, which represents the signal as a sum of infinite periodic harmonics, is not naturally suited for decomposition of short and aperiodic percussion signals. Broadening of the spectral peaks due to damping leads to their overlapping and masking of the lower amplitude peaks, which could be important for the fine-level signal classification. In this study, an attempt is made to automatically decompose percussion signals into a sum of exponentially damped harmonics, which in this case form a more natural basis than Fourier harmonics and thus allow for a more robust representation of the signal in the parametric space. The damped harmonic decomposition of percussion signals recorded on healthy volunteers in clinical setting is performed using the matrix pencil method, which proves to be quite robust in the presence of noise and well suited for the task.

Estimating the parameters of audible clinical percussion signals by fitting exponentially damped harmonics

Used for centuries in the clinical practice, audible percussion is a method of eliciting sounds by tapping various areas of the human body either by finger tips or by a percussion hammer. Despite its advantages, pulmonary diagnostics by percussion is still highly subjective, depends on the physician's skills, and requires quiet surroundings. Automation of this well-established technique could help amplify its existing merits while removing the above drawbacks. In this work, clinical percussion signals from normal volunteers are decomposed into a sum of exponentially damped sinusoids (EDS) whose parameters are determined using the Matrix Pencil Method. Some EDS represent transient oscillation modes of the thorax/abdomen excited by the percussion event, while others are associated with the noise. It is demonstrated that relatively few EDS are usually enough to accurately reconstruct the original signal. It is shown that combining the frequency and damping parameters of these most significant EDS allows for efficient classification of percussion signals into the two main types historically known as "resonant" and "tympanic." This classification ability can provide a basis for the automated objective diagnostics of various pulmonary pathologies including pneumothorax. The algorithm can be implemented on an embedded platform for the battlefield and other emergency applications.