Fig 1.
Schematic of the experimental setup and representative dataset.
The animal model is anesthetized and attached to a medical monitoring system that measures ECG, capnography, and photoplethysmography (PPG) waveforms simultaneously. We built a custom voltage level-shifting circuit for each waveform, which outputs to a commercial A/D converter. Two electret microphones, one controlled internally by the endoscope, the other attached superficially on the pig’s chest just above the heart, are also sending data to the A/D converter. The final result is perfectly time-registered data streams for heart and lung function as well as the acoustic waveforms. Example concurrent physiological data measurements were taken from the proximal third of a porcine stomach including the acoustic waveform from our internal electret microphone, ECG, PPG (which indicates systemic oxygen perfusion levels and heart rate), capnography from expired CO2 content, and the acoustic waveform from the external microphone positioned above the heart. (Note that the raw PPG data from the SurgiVet system seems to be inverted, and the raw capnography data appears to be the flow rate of CO2, thus giving a first derivative of the more familiar capnogram waveform.).
Fig 2.
Schematic processing flow chart for HR and RR estimation from internal microphone data.
The signal is copied into a HR track and RR track and then analog filtered and down-sampled. A sliding window computes an energy feature (see Methods) that is input to the average magnitude difference function (AMDF). The RR is further low pass filtered with a 1 second Hamming window. The first valley of the AMDF is the estimated vital sign.
Fig 3.
Example spectrogram and corresponding time course data of the internal acoustic signal measured in the proximal third of the stomach.
A majority of the signal energy is <50Hz.
Fig 4.
HR and RR estimation performance histograms for all data collected as a function of anatomical location.
The x-axis represents the absolute value of the percent error in a given 20s frame; the y-axis is the number of such frames normalized by the “Total Counts” (for each location and vital sign) used to build the histogram. The AMDF valley-finding algorithm can trigger upon higher-order harmonics of the fundamental period, on noise, or on the incompletely removed heart rate period, giving percent errors concentrated at 50% and 100%; these errors can be easily addressed with more sophisticated AMDF algorithms or running median filters (see Discussion).
Fig 5.
Vital sign algorithm performance as a function of anatomical site.
Median absolute percentage error between PPG and capnography derived HR and RR and acoustically derived HR and RR via our average magnitude difference function (AMDF)-based analysis is reported at each measurement site.