Body sounds are one of the oldest physiological parameters known to be used in diagnostics. Even today the stereotypical image of a physician is one of a person with a stethoscope around the neck. Indeed body sounds are still today a very important parameters that can relay a lot of disease information. Heart rate and abnormal respiration patterns are clear and well known examples. However today body sounds are still primarily recorded via (digitally amplified) stethoscopes and analysis is done by the physician. A clear trend in healthcare is towards more and more remote and automatic monitoring. A lot of effort has been spent already on miniaturizing and automating bio-potential recording (for example electro-cardiogram or ECG). However, very little research effort has been conducted towards this goal on body sounds. On the other hand, the rapid advancements in machine learning have enables a whole range of innovations for speech recognition. Technological advancements in low-power, low-cost integrated electronics as well as artificial intelligence have paved the way for ultra-low-power always-on speech recognition processors that can detect a few words at uW power levels.
In this PhD, we will explore the possibility to apply similar strategies to detect abnormal respiration patterns in recorded body sounds to classify for example certain specific respiratory diseases. The scientific hypothesis is that certain analysis of body sounds can be considered as a pattern recognition and classification problem, conceptually similar to speech recognition. The candidate will be asked to first explore suitable algorithms. This will also require investigation into miniaturized recording of body sounds in a wearable form factor since it is to be expected that the recordings will be of significantly lower quality (ambient noise pickup, artifacts from motion, ...) than high-fidelity audio recordings with a digital stethoscope. The candidate will use these learnings combined with high-quality recordings from publicly available databases to develop as suitable large database. This database will then be used to develop and validate machine-learning algorithms. Finally the candidate will develop an ultra-low-power body-sounds processor in an ASIC and will validate the work in an experimental setting. The algorithms and processor will be co-optimized to come to the most efficient end-to-end system solution.
Figure 1: Prototype of a wearble patch with an embedded body sounds microphone.
Required background: electrical engineering with experience in digital implementations, computer architectures and machine learning
Type of work: machine learning based algorithm development (40%), digital IC design implementation (40%), testing and validation (20%)
Supervisor: Marian Verhelst
Daily advisor: Dwaipayan Biswas
The reference code for this position is 2020-107. Mention this reference code on your application form.
Chinese nationals who wish to apply for the CSC scholarship, should use the following code when applying for this topic: CSC2020-56.