Type of study: An analytic experimental study in an ex-vivo porcine model.
Study design: The study was conducted in four white domestic female pigs (Laboratory Animals Farm, Lahav, Israel), aged 4 months (41–50 kg), housed in the institutional animal facility accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC). Water and normal appropriate diet were available ad libitum. The experimental procedures were performed seven days after acclimatization. Food was withheld from the night before the procedure. Animals were sedated with Xylazine (1 mg/kg, IM, Eurovet Animal Health BV, The Netherlands) and anesthesia was induced with Ketamine (10 mg/kg, IM, Vetoquinol SA, Lure Cedex, France). The ear vein was then cannulated for intravenous administration of a mixture of Diazepam (2 mg, IV, TEVA Pharmaceutical Industries Ltd, Noida, India), Ketamine (400 mg, IV), Propofol (1–4 mg/kg, IV, Fresenius Kabi Austria Gmbh, Austria), and Tramadol (5 mg/kg, IM, Rafa Laboratories Ltd, Jerusalem, Israel) for analgesia. The pigs were then intubated with a cuffed silastic endotracheal tube (7.0-mm, Portex Tracheal Tube, UK). Anesthesia was maintained with 2% isoflurane (Piramal Critical Care Inc, Pa) in 100% oxygen, and animals were ventilated using controlled mechanical ventilation (Excel 210-SE anesthesia machine Datex-Ohmeda Inc, Madison, Wis) or Narkomed-2B Anesthesia Machine - North American Drager, Pa). Tidal volume was set to 10 mL/kg with respiratory rate of 13 to 15 breaths per minute adjusted to an end-tidal CO2 (ETCO2) 35 mm Hg at baseline. The animals were euthanized with an intravenous injection of KCl solution (Fagron Group BV, Rotterdam, The Netherlands). The porcine model was chosen due to its physiological and anatomical similarity to the human thorax including chest wall thickness, subcutaneous fat distribution, skin thickness, intercostal spaces width, muscle tissue response, and lung positioning. Moreover, The volumes of porcine and humans’ lungs are considered comparable, with a total lung capacity of 55 mL/kg .
Twelve seconds recordings of breath sounds, representing three mechanical inhalation-exhalation cycles, were sampled over the right 7th intercostal space, at the mid-axillary line. This specific site was chosen based on previous studies performed by our group as an appropriate site for auscultation in similar-size animals. Preliminary baseline recordings were sampled and were defined as "volume 0". After this, simple PTX was induced with increasing volumes of 200, 400, 600, 800, and 1,000 ml of air, injected using a 100 ml syringe and a 14G needle into the pleural cavity and underneath the sampling site. Before simulating HTX, the thorax was emptied of air by using the same syringe, until air leak through the needle was absent. Then, an increasing volume of saline (0.9% sodium chloride) was injected in the same position as described above: 200, 400, 600, 800, and 1,000 ml. Each of the aforementioned states was sampled for a twelve-second recording of breath sounds. The study was conducted in a quiet surrounding, without restrictions on environmental noise during the recordings.
a. Apparatus assembly: We chose to use two FDA- and CE-approved commercial Thinklabs One digital stethoscopes (Thinklabs Medical LLC, CO, USA) , capable of sampling a wide range of frequencies and amplifying them by up to 100-fold, allowing for recording over a thick chest wall if needed. The stethoscope is round-shaped, with a diameter of 46 mm and a height of 28 mm, it weighs about 50 grams and is equipped with built-in filters that reduce background noise. The filter we chose was designed to enable a sampling of 20-2,000 Hz +/-3 dB. We placed two microphones next to each other, supported by a longitudinal sustaining frame. The microphones were connected to a spatial sampling system assembled at the IDF Medical Corps Medical Engineering Division. The system included a 3.5 mm adaptor connecting the stethoscopes to a portable myRIO-1900 embedded device (National Instruments, TX, USA) . The sampling apparatus is presented in Figure 2.
b. Software programming: The sampling algorithm was programmed by the IDF Medical Corps Medical Engineering Division using the LabVIEW platform. The samples recorded by the two stethoscopes were separated into two audio channels for each stethoscope. Therefore, for each recording made, four data sources were obtained. The recordings were analyzed using a MATLAB computing environment and were filtered to omit high-pass and low-pass data points with frequencies of <60 Hz and >2,000 Hz, respectively, that were not originated from the respiratory system. For each recording, the intensity ( ) of the signal was measured in constant gaps of 0.00005 seconds, i.e. 20,000 samples per second. In total, 176 datasets were obtained and were later extracted to an Excel sheet for further data processing and analysis by an external software house (Clinetix - Clinical Kinetics, Tel-Aviv, Israel). The working assumption was that each volume of simple PTX and HTX had a unique acoustic pattern that could be obtained and analyzed, and on which a classification algorithm could be programmed and enable the differentiation of the various volumes of injury compared to the normal state. Due to the small database, we decided not to use a model-based analytical method, such as an artificial neuronal network or a decision tree, which is a better fit for larger quantities of datasets. This led us to use a memory-based method, namely a k-nearest neighbors (KNN) algorithm, which generates prediction by estimating how likely a data point is to be a member of one group or another, depending on what group the data point nearest to it is in. A multi-objective genetic algorithm (MOGA) was used to develop a mathematical detector through the process of artificial evolution. The detector and its features, which are a set of mathematical formulas, were developed by the MOGA as also the KNN algorithm’s parameters.
Statistical analysis: Sensitivity, specificity, and accuracy for each injury volume were calculated. Accuracy was defined as
(TP-True positive, TN-True negative, FP-False positive, FN-False negative). In addition to comparing each of the volumes separately from each other, we conducted additional testing combining all observations with comparison to baseline (“volume 0”). A Pearson correlation coefficient was computed to assess the relationship between the real value and the predicted one. Calculations were made using a C# programming language in a LINQPad software utility.
Ethics: This study conformed to the Guide for the Care and Use of Laboratory Animals, (National Academy Press, Washington, DC, 1996). Animal care and experimental procedures were approved by the Ethics Committee of The Faculty of Medicine of The Hebrew University of Jerusalem, Israel (approval numbers MD-16-14853-3 and MD-17-15043-3).
Funds: The study was conducted with the help of a research grant from the IDF Medical Corps, and from the Director of the Administration for the Development of Arms and Technological Infrastructure (Maf'at) at the Israeli Ministry of Defense (grant number: 4440838636).