The research question in this thesis concerns how well can epileptic seizures be detected using a single triaxial accelerometer attached to the wrist. This work was done in collaboration with Vivago Oy, who provided the watch that is capable of recording accelerometer data, and HUS, the Hospital District of Helsinki and Uusimaa. HUS provided the real world epilepsy datasets consisting of several days worth of data recorded by several epilepsy patients.
The research problem was divided into three subproblems: feature extraction, sensor fusion, and activity classification. For feature extraction, the original accelerometer signal is divided into 5s long windows and discrete cosine transform (DCT) is applied to each axis so that periodic components are detected, also removing the effect of gravity vector and compressing the signal. Next, the DCT features of each axis are combined and principal component analysis (PCA) is applied, further compressing the signal. At this step the PCA theorem is also proven. After DCT and PCA steps, the need to consider for different orientations of the accelerometer is effectively eliminated. The last step is the classification of the signal into a seizure or non-seizure by using a support vector machine (SVM) classifier on the features produced by PCA. The combined model is referred to as the DPS model (DCT-PCA-SVM).
The experiments were run on two kinds of datasets: artificial datasets recorded by three test subjects and the epilepsy datasets. The principal reason for recording artificial datasets was that the labeling of the seizures in the epilepsy dataset was practically impossible to match to the accelerometer data, rendering the supervised training phase for any model impossible. The artificial datasets were created so that one test subject produced the training set, recording data of ordinary daily activities and labeling these activities as non seizures, and then imitating a seizure and labeling this as a seizure. The second test subject recorded the daily activities, including potential false positives such as brushing teeth and washing hands, and imitating a seizure several times during this period. This validation set was then used for fine-tuning the DPS model parameters so that all of the seizures were detected along with as few false positives as possible. Third test subject recorded the test set, including 13 imitated seizures, to test the DPS model's ability to generalize on new and previously unseen data.
The conclusion is that for the artificial test set, 12 out of 13, or 92%, of seizures were detected along with a reasonably low number of false positives. For the epilepsy dataset the results are inconclusive, due to not being able to utilize any part of it as a training set, but there are reasonable indications that at least some real seizures were detected. In order to verify the results, the DPS model would need to be trained on a larger and better labeled real world epilepsy dataset.
