Voxel Weight Matrix-Based Feature Extraction for Biomedical Applications

Fahad Albalawi, Sultan Alshehri, Abderrazak Chahid, Taous-Meriem Laleg-Kirati

Research output: Contribution to journalArticlepeer-review

Abstract

functional Magnetic Resonance Imaging (fMRI) is an emerging medical tool used to measure brain activities that were induced normally such as cognitive states (e.g., reading a sentence or viewing a picture) or abnormally (e.g., brain activity occurs after a stroke or brain injury). These measured data can be used to construct a model via machine learning techniques to predict the occurrence of a certain cognitive behavior or brain disease. The difficulty of this prediction problem can be summarized in two points: first, the size of the dataset is very small due to the small number of subjects (i.e., patients) who can contribute to these research-based experiments. Second, the size of the feature vector resulted from these medical tools is very large compared to the few number of samples that were collected. One possible way to overcome these obstacles is to develop a feature generation methodology that can produce a small-sized and descriptive feature vector that may improve the overall prediction performance. Motivated by these considerations, this paper proposes a novel feature generation methodology termed Voxel Weight Matrix (VWM) method. This feature generation technique can transform the original high-dimensional feature vector to a two-dimensional discriminative feature domain. The main contribution of this feature generation technique is its ability to represent the statistical measures of the original feature vector via two-dimensional feature vector. After generating the VWM-based feature set, various classification tools such as logistic regression (LR) models and Support Vector Machine (SVM) are used for cognitive state prediction based on publicly available fMRI dataset called state/plus dataset. The classification models with the proposed VWM features outperformed the best two reported prediction models associated with the star/plus dataset with an average accuracy of 99.8%. To further illustrate the effectiveness of the proposed feature generation methodology, another publicly available Electroencephalography (EEG) dataset are used for Epileptic Seizure Prediction.
Original languageEnglish (US)
Pages (from-to)1-1
Number of pages1
JournalIEEE Access
DOIs
StatePublished - 2020

Fingerprint

Dive into the research topics of 'Voxel Weight Matrix-Based Feature Extraction for Biomedical Applications'. Together they form a unique fingerprint.

Cite this