Repository for bachelor thesis on Automatic Multi-Modal Detection of Autonomic Arousals in Sleep. The thesis itself (236 pages) and all related data is confidential and thus not publicly available, but access to the thesis can be granted upon request. We have however made our presentation slides available for review here.

Manual scoring of arousals is a tiring process prone to high inter-scorerdisagreement. Current arousal detection is based on electroencephalog-raphy (EEG), which is not suitable for sleep studies at home. Basner et al. (2008) proposes a semi-automatic detection algorithm based on electrocardiography (ECG) and acknowledges that it cannot replace EEG, but a strong correlation between cortical and autonomic arousals is evident. This study explores the possibility of better autonomic arousal detection algorithms in a multi-modal setting by inclusion of features from photoplethysmography (PPG).

A state-of-the-art algorithm provided by M. Olsen (2018) is applied for reliable RR-tachogram extraction in ECG signals, and a pulse peak detection algorithm is developed for extraction of pulse transit time and corresponding pulse wave amplitude in PPG signals. Classification is performed by a recurrent neural network (RNN) utilising gated recurrent unit (GRU) cells with an architecture derived from model- and feature selection using 2-layered cross-validation on the Multi-Ethnic Study of Atherosclerosis (MESA) data-set.

It was found that the best performing model was a 2-layered bi-directional RNN using GRU cells, and cross-validation shows slight performance increase when introducing PTT and PWA to a RR-interval based model. The best performing modality in this study was an RR-interval calculated at 800Hz upsampling. A generalisation estimation is performed on a sub-set of an additional data-set, Sleep Heart Health Study (SHHS), in order to provide unbiased results for comparison to other studies. A sensitivity of 58.53%, precision of 71.68%, F1-score of 0.6444 and a correlation between scored and predicted arousal indices of (R=0.762, p<0.01) was found. These results compare well to state-of-the-art research but with a use of far less features. It was found that the model performed poorly on subjects with abnormal heart rates, suggesting ECG-based algorithms cannot reliably be employed unless deployed multi-modally.

A few figures from the thesis illustrating our work and results is shown below.

You are more than welcome to contact me directly at hello@nicklashansen.com if you have any questions regarding our work.