Development of a graphical user interface for automatic separation of human voice from Doppler ultrasound audio in diving experiments.

Doppler ultrasound (DU) is used in decompression research to detect venous gas emboli in the precordium or subclavian vein, as a marker of decompression stress. This is of relevance to scuba divers, compressed air workers and astronauts to prevent decompression sickness (DCS) that can be caused by these bubbles upon or after a sudden reduction in ambient pressure. Doppler ultrasound data is graded by expert raters on the Kisman-Masurel or Spencer scales that are associated to DCS risk. Meta-analyses, as well as efforts to computer-automate DU grading, both necessitate access to large databases of well-curated and graded data. Leveraging previously collected data is especially important due to the difficulty of repeating large-scale extreme military pressure exposures that were conducted in the 70-90s in austere environments. Historically, DU data (Non-speech) were often captured on cassettes in one-channel audio with superimposed human speech describing the experiment (Speech). Digitizing and separating these audio files is currently a lengthy, manual task. In this paper, we develop a graphical user interface (GUI) to perform automatic speech recognition and aid in Non-speech and Speech separation. This constitutes the first study incorporating speech processing technology in the field of diving research. If successful, it has the potential to significantly accelerate the reuse of previously-acquired datasets. The recognition task incorporates the Google speech recognizer to detect the presence of human voice activity together with corresponding timestamps. The detected human speech is then separated from the audio Doppler ultrasound within the developed GUI. Several experiments were conducted on recently digitized audio Doppler recordings to corroborate the effectiveness of the developed GUI in recognition and separations tasks, and these are compared to manual labels for Speech timestamps. The following metrics are used to evaluate performance: the average absolute differences between the reference and detected Speech starting points, as well as the percentage of detected Speech over the total duration of the reference Speech. Results have shown the efficacy of the developed GUI in Speech/Non-speech component separation.

An open-source framework for synthetic post-dive Doppler ultrasound audio generation.

Doppler ultrasound (DU) measurements are used to detect and evaluate venous gas emboli (VGE) formed after decompression. Automated methodologies for assessing VGE presence using signal processing have been developed on varying real-world datasets of limited size and without ground truth values preventing objective evaluation. We develop and report a method to generate synthetic post-dive data using DU signals collected in both precordium and subclavian vein with varying degrees of bubbling matching field-standard grading metrics. This method is adaptable, modifiable, and reproducible, allowing for researchers to tune the produced dataset for their desired purpose. We provide the baseline Doppler recordings and code required to generate synthetic data for researchers to reproduce our work and improve upon it. We also provide a set of pre-made synthetic post-dive DU data spanning six scenarios representing the Spencer and Kisman-Masurel (KM) grading scales as well as precordial and subclavian DU recordings. By providing a method for synthetic post-dive DU data generation, we aim to improve and accelerate the development of signal processing techniques for VGE analysis in Doppler ultrasound.

A fully automated algorithm for heart rate detection in post-dive precordial Doppler ultrasound.

Background: Doppler ultrasound is used currently in decompression research for the evaluation of venous gas emboli (VGE). Estimation of heart rate from post-dive Doppler ultrasound recordings can provide a tool for the evaluation of physiological changes from decompression stress, as well as aid in the development of automated VGE detection algorithms that relate VGE presence to cardiac activity. Method: An algorithm based on short-term autocorrelation was developed in MATLAB to estimate the heart rate in post-dive precordial Doppler ultrasound. The algorithm was evaluated on 21 previously acquired and labeled precordial recordings spanning Kisman-Masurel (KM) codes of 111-444 (KM I-IV) with manually derived instantaneous heart rates. Results: A window size of at least two seconds was necessary for robust and accurate instantaneous heart rate estimation with a mean error of 1.56 ± 7.10 bpm. Larger window sizes improved the algorithm performance, at the cost of beat-to-beat heart rate estimates. We also found that our algorithm provides good results for low KM grade Doppler recordings with and without flexion, and high KM grades without flexion. High KM grades observed after movement produced the greatest mean absolute error of 6.12 ± 8.40 bpm. Conclusion: We have developed a fully automated algorithm for the estimation of heart rate in post-dive precordial Doppler ultrasound recordings.

Deep Learning-Based Venous Gas Emboli Grade Classification in Doppler Ultrasound Audio Recordings.

OBJECTIVE: Doppler ultrasound (DU) is used to detect venous gas emboli (VGE) post dive as a marker of decompression stress for diving physiology research as well as new decompression procedure validation to minimize decompression sickness risk. In this article, we propose the first deep learning model for VGE grading in DU audio recordings. METHODS: A database of real-world data was assembled and labeled for the purpose of developing the algorithm, totaling 274 recordings comprising both subclavian and precordial measurements. Synthetic data was also generated by acquiring baseline DU signals from human volunteers and superimposing laboratory-acquired DU signals of bubbles flowing in a tissue mimicking material. A novel squeeze-and-excitation deep learning model was designed to effectively classify recordings on the 5-class Spencer scoring system used by trained human raters. RESULTS: On the real-data test set, we show that synthetic data pretraining achieves average ordinal accuracy of 84.9% for precordial and 90.4% for subclavian DU which is a 24.6% and 26.2% increase over training with real-data and time-series augmentation only. The weighted kappa coefficients of agreement between the model and human ground truth were 0.74 and 0.69 for precordial and subclavian respectively, indicating substantial agreement similar to human inter-rater agreement for this type of data. CONCLUSION: The present work demonstrates the first application of deep-learning for DU VGE grading using a combination of synthetic and real-world data. SIGNIFICANCE: The proposed method can contribute to accelerating DU analysis for decompression research.

A Review of Deep Learning-Based Contactless Heart Rate Measurement Methods.

The interest in contactless or remote heart rate measurement has been steadily growing in healthcare and sports applications. Contactless methods involve the utilization of a video camera and image processing algorithms. Recently, deep learning methods have been used to improve the performance of conventional contactless methods for heart rate measurement. After providing a review of the related literature, a comparison of the deep learning methods whose codes are publicly available is conducted in this paper. The public domain UBFC dataset is used to compare the performance of these deep learning methods for heart rate measurement. The results obtained show that the deep learning method PhysNet generates the best heart rate measurement outcome among these methods, with a mean absolute error value of 2.57 beats per minute and a mean square error value of 7.56 beats per minute.