Engineering and Design
From Concept to Completed Prototype
Design Inputs
Through interviews with physicians, searches of the research literature, and many brainstorming sessions, we identified several functional requirements, constraints, and customer needs, all of which can be summarized in three primary design inputs:
1. Increase the sensitivity and specificity of auscultation
2. Identify regions with abnormal sounds
3. Integrate with current clinical practice
User Experience
Our final prototype balances ease of use and accuracy of placement by integrating ten transducers into a single system of straps. The patient puts the straps on themselves and tightens them as necessary, then the doctor places adhesive diaphragms at the correct anatomical locations. The diaphragms attach to the microphones with snap-in connectors, and the entire set up can be completed in under a minute and a half.
Once the device is in position, the doctor simply instructs the patient to breath and records for any length of time through our intuitive user interface. After the recording is complete, the results fill the diagram of the lung. Each number is a normalized intensity value and each of the four colors expresses how confident the software is that the sound is normal or adventitious. Clicking on each location allows for playback of the sound, and all the data can be saved for longitudinal comparison.
How It's Done
In order to make this happen, we performed some analog filtering and amplification to clean up the sounds and built a software classification program to identify normal and adventitious sounds. Through frequency analysis of a database of over 100 pre-recorded lung sounds, we identified two main features that could be used to classify normal and adventitious sounds. The first feature was a ratio of the energy in low versus high frequencies, and the second feature was based on the maximum energy in the 500-700 Hz range. Together these features formed two-dimensional classifier shown in the graph to the right.
Testing Outcomes
The classifier was validated using a leave-m-out testing protocol, and the sensitivity and specificity of our program were found to be consistently above 90%.
The ability of our device to localize regions of abnormal sound was tested by selectively playing a sound with a frequency characteristic of adventitious sounds at one location, while playing sounds with frequencies characteristic of normal sounds everywhere else. The success of this test proved that the device functions properly as a whole - from the wiring of the microphones and circuitry, to the analysis algorithms written in the code, to the visual output designed in the user interface.
Our final test was an assessment of the form factor of our device, or in other words, how easy it is to wear and how accurately it positions the transducers on patients of different sizes. We had several patients of varying heights, weights and genders try on our device and successfully concluded that the device could be easily adjusted to place the transducers at the correct locations across the lungs.
