【博士論文】学術データベース

博士論文 / Classification of Epileptic Seizure EEG Signals in Time Frequency Domain - Focusing on Root Mean Square based Feature Extraction 時間周波数領域でのてんかん脳波識別に関する研究 ‐平均二乗根に基づく特徴抽出に着目して‐

著者

書誌事項

タイトル

Classification of Epileptic Seizure EEG Signals in Time Frequency Domain - Focusing on Root Mean Square based Feature Extraction

タイトル別名

時間周波数領域でのてんかん脳波識別に関する研究 ‐平均二乗根に基づく特徴抽出に着目して‐

著者名

Mahapatra Arindam Gajendra

学位授与大学

九州工業大学 (大学ID:0071) (CAT機関ID:KI000844)

取得学位

博士(工学)

学位授与番号

甲生工第323号

学位授与年月日

2018-06-28

注記・抄録

Epilepsy affects over 50 million people on an average yearly world wide. Epileptic Seizure is a generalised term which has broad classification depending on the reasons behind its occurrence. Parvez et al. when applied feature instantaneous bandwidth B2AM and time averaged bandwidth B2FM for classification of interictal and ictal on Freiburg data base, the result dipped low to 77.90% for frontal lobe whereas it was 80.20% for temporal lobe compare to the 98.50% of classification accuracy achieved on Bonn dataset with same feature for classification of ictal against interictal. We found reasons behind such low results are, first Parvez et al. has used first IMF of EMD for feature computation which mostly noised induce. Secondly, they used same kernel parameters of SVM as Bajaj et al. which they must have optimised with different dataset. But the most important reason we found is that two signals s1 and s2 can have same instantaneous bandwidth. Therefore, the motivation of the dissertation is to address the drawback of feature instantaneous bandwidth by new feature with objective of achieving comparable classification accuracy. In this work, we have classified ictal from healthy nonseizure interictal successfully first by using RMS frequency and another feature from Hilbert marginal spectrum then with its parameters ratio. RMS frequency is the square root of sum of square bandwidth and square of center frequency. Its contributing parameters ratio is ratio of center frequency square to square bandwidth. We have also used dominant frequency and its parameters ratio for the same purpose. Dominant frequency have same physical relevance as RMS frequency but different by definition, i.e. square root of sum of square of instantaneous band- width and square of instantaneous frequency. Third feature that we have used is by exploiting the equivalence of RMS frequency and dominant frequency (DF) to define root mean instantaneous frequency square (RMIFS) as square root of sum of time averaged bandwidth square and center frequency square. These features are average measures which shows good discrimination power in classifying ictal from interictal using SVM. These features, fr and fd also have an advantage of overcoming the draw back of square bandwidth and instantaneous bandwidth. RMS frequency that we have used in this work is different from generic root mean square analysis. We have used an adaptive thresholding algorithm to address the issue of false positive. It was able to increase the specificity by average of 5.9% on average consequently increasing the accuracy. Then we have applied morphological component analysis (MCA) with the fractional contribution of dominant frequency and other rest of the features like band- width parameter’s contribution and RMIFS frequency and its parameters and their ratio. With the results from proposed features, we validated our claim to overcome the drawback of instantaneous bandwidth and square bandwidth.

九州工業大学博士学位論文 学位記番号:生工博甲第323号 学位授与年月日:平成30年6月28日

1 Introduction|2 Empirical Mode Decomposition|3 Root Mean Square Frequency|4 Root Mean Instantaneous Frequency Square|5 Morphological Component Analysis|6 Conclusion

平成30年度

九州工業大学博士学位論文(要旨)学位記番号:生工博甲第323号 学位授与年月日:平成30年6月28日

キーワード

SVM, EEG, EMD, MCA, RMS Frequency, RMIFS Frequency

各種コード

NII論文ID(NAID)

500001323539

NII著者ID(NRID)
  • 8000001440375
本文言語コード

eng

データ提供元

機関リポジトリ / NDLデジタルコレクション

博士論文 / 九州工業大学 / 工学

博士論文 / 九州工業大学

博士論文 / 工学

関連著者

博士論文 / 大学

博士論文 / 学位