Spark and Rule-KNN based scalable machine learning framework for EEG deceit identification. (April 2020)

• Record Type: Journal Article
• Authors: Thakur, Santosh; Dharavath, Ramesh; Edla, Damodar Reddy
• Is Part Of: Biomedical signal processing and control. Volume 58(2020)
• Publication Date: 2020-04
• Subjects: Apache spark -- Non-parametric -- Weighted feature extraction -- EEG data analysis -- Brain computer interface; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2020.101886
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel framework for EEG based Deceit identification using Apache Spark Framework is presented. The non-parametric weighted feature extraction (NPWFE) is applied to extract the features from EEG signals. KNN is applied to classify the feature groups. A rule set was constructed by ranking the discrimination rules according to the kappa agreement. The experiment is conducted on 20 healthy subjects and collected real time EEG data. Abstract: Brain computer interface (BCI) provides communication between the computer and the brain. It is the combination of hardware and software which provides non-muscular channel to send the various messages to control the computer. BCI is useful in various medical applications such as patients with neuromuscular injuries, locked-in syndrome (LiS) etc. BCI is not only useful in medical applications, but also useful in lie detection, entertainment, etc. In this paper, spark and rule-KNN based scalable framework has been presented using BCI with the EEG data collected on 20 subjects in which 10 are acted as innocent and 10 are acted as guilty. Using BCI P300, Deceit identification Test (DIT) is performed. To perform DIT, we classify the P300 signals which have a positive peak of 300 ms–1000 ms in one stimulus start. Data processing is performed with band pass filter to cut the frequency ranges and features are extracted using non-parametric weighted feature extraction followed by rule based discriminant classification. For training and testing, the data ratio selected as 80:20 and achieved the accuracy 92.46 %. Proposed framework provides better results in comparison with existing models presented in literature. Hence this model is accurate, scalable and fault tolerant.