Improving Data quality for the detection of Obstructive Sleep Apnea(OSA), using adversarial training for deep neural architectures
Abstract: In many real-world situations, biological data collected from body sensors have low data quality due to various reasons (misplacement, bad sensor quality, movement of the patient etc). This can negatively impact the potential application of the training procedures of machine learning/ deep learning models, and thus decrease their class decision performance. Real world OSA recordings are a prime example of this type of data . Many signal processing techniques exist that preprocess the data before the classification procedure in order to make them more suitable for training. Adversarial training is an idea mostly used in the context of deep learning, where an optimization algorithm tries to find and exploit weaknesses in the deep learning model by slightly altering the data. However with this procedure the deep model also learns the manipulation and thus becomes better at handling it. We are interested in employing this idea to low quality real world OSA recordings, where the manipulation of the adversary will be related to a mapping of the low quality behavior.
Requirements:
-Basic knowledge of Python
-Basic calculus and probability theory background
-Interest in machine learning
Contact/supervisors:
Veileder(e)
Visualization of training for time series data
Abstract: Recently there have been attempts at understanding the training procedure of deep networks but most have been focusing on images, especially for convolutional neural networks. In a similar fashion we are interested in using visualization to get a better understanding of the training for time series data, using mostly recurrent neural nets.
Requirements:
-Basic knowledge of Python
-Basic calculus and probability theory background
-Interest in machine learning
Contact/supervisors:
- Konstantinos Nikolaidis
- Stein Kristiansen
- Vera Hermine Goebel
Extending the range of Bluetooth LE for reliable wireless monitoring of physiological signals
In the Cesar project we are using several different technologies to monitor sleep of individuals at home with smartphones that are connected via Bluetooth LE to on body sensors, like the Flow sensor from Sweetzpot (https://www.sweetzpot.com/flow). It has turned out that during overnight monitoring the connection between smartphone and sensor gets broken. One reason for the connection loss is that the human body is in between the sensor and the smart phone. During the period without connectivity data gets lost. To enable reliable sleep monitoring we aim to investigate in the thesis the use of Bluetooth LE repeaters and other technologies.
Useful knowledge:
- Java
- Android
- Bluetooth
Contact/supervisor: