Subtitle: Deep learning methods for improving essential hypertension prediction based on magnetoencephalography data
Supervisors: Justin Schneiderman (Chalmers, MC2), Kevin Andersson (Syntronic Research and Development AB)
Opponents: Bastian Jung and Rami Salem
Examiner: Paolo Monti (E2)
Cardiovascular disease is a wide-spread problem which is strongly linked to essential hypertension (high blood pressure). Previous research has demonstrated a correlation between essential hypertension and muscle-sympathetic nervous activity (MSNA), which can be measured invasively. Research has also shown that there is a correlation between MSNA and MEG brain activity during stressful situations. Two previous Master’s thesis projects have tried to make a classifier from this data with limited success.
This Master’s thesis project aims to improve classification results by resolving any issues in previous theses and by trying new approaches to the problem. The main hypotheses investigated are 1) Inappropriate data split, 2) Lack of data augmentation, 3) Inappropriate model choice. To test these hypotheses, a more principled data split was used, in combination with cross-validation. Furthermore, several augmentation methods and deep learning architectures have been implemented and optimized for using Bayesian hyperparameter tuning. This resulted in test accuracies higher than those reported in previous projects, but with strong indications that the performance is still largely explained by random chance. In conclusion, it seems that the small dataset size is still the major reason for the limited success, and future work in this direction should strongly consider incorporating additional data sources for the analysis.
David Hall and Elias Sundqvist
E2 Room 7430 Landahlsrummet