Lachit Dutta

Two-fold feature extraction technique for biomedical signals classification

This presents a two-fold feature extraction technique for biomedical signals classification. In this work, signals are uniformly decomposed to form a set of uniform matrices. Then using Canonical Correlation Analysis (CCA) each pair matrices are mapped to orthogonal space. Next, Wavelet transformation is performed on original feature matrices and then mapped to orthogonal space. Both domains are statistically independent. From each domain, same dimensional feature vectors are extracted and concatenated them to form single embedding vectors. The embedding vectors are fed to classifier to recognize the healthy control and pathological signal patterns. For demonstration, we consider three groups of EMG signal vis Amyotrophic lateral sclerosis (ALS), Myopathy (Myo) and healthy control (Nor). Results indicate that adopted feature extraction technique and synchronization of features strongly enhances the quality of feature pattern. The optimum recognition rate under adopted feature technique are obtained 95.91%±3.6% and 95.58%±1.5 % in Myo-Nor and ALS-Nor respectively. The proposed feature extraction scheme is consistent not only in accuracies but also other quality assessment parameters. Hence it promises to provide a better strategic tool for signal classification.

Comparison of direct interfacing and ADC based system for gas identification using E-Nose

This paper presents the design and characterization of a direct interfacing circuit (DIC) for sensor array and its comparison with the built-in analog to digital converter (ADC) of a PIC microcontroller. Before implementing the DIC for the sensor array, we examined its performance on simulated voltages to have a proper understanding of the output characteristics. To explore the DIC for multi-sensory system, an exemplary E-Nose setup was developed for experimentation. The sensor responses from the E-Nose system are concurrently measured by two microcontrollers, one using DIC and the other by the 10-bit internal ADC of the microcontroller. Principal component analysis (PCA) is then performed for visualizing and comparing the class separation for both the methods. To further explore the discriminating capability of the DIC based E-Nose, artificial neural network (ANN) is implemented. Finally, we delve into microcontroller based online gas discrimination by the E-Nose using DIC and its comparison with ADC results. Equivalent results were observed for both the cases with accuracy up to 97 %.

Fusion of projected feature for classification of EMG patterns

In medical domain, proper set of information fusion is the vital requirement for quantitative interpretation. In this paper, we present a novel feature fusion technique based on two-fold feature projection (FP) for EMG classification. Canonical Correlation Analysis (CCA) transformation is performed on original feature space and wavelet transformation of original feature. Based on subspace learning technique, relevant features are extracted from unified domain independent spaces and fused via proposed fusion algorithms. To demonstrate the outperform of the adopted algorithm three class of EMG subject groups are considered from online and Guwahati Neurological Research Centre (GNRC), Ghy, India. Outcomes indicate that the adopted dimension reduction strategy are consistent not only in accuracy but also in other quality assessment parameters. The overall accuracy is 98.80% ±2.0%. It provides an efficient and powerful way of feature extraction for fusion to improve recognition rate. Hence, it promises to prove a better strategic tool for medical data analysis in healthcare institutions.

F-SVD based algorithm for variability and stability measurement of bio-signals, feature extraction and fusion for pattern recognition

Abstract A support system with efficient learning framework helps eliciting complete knowledge of underlying phenomena of interest. It makes the analysis less-onerous, time-consuming and error-prone and thus promotes large scale applications. Such modeling requires profound understanding of available information and its appropriate utilization. Albeit success of electromyogram (EMG) support systems, challenges still exits specifically in early phase of design mainly due to inherent variations and complex data distribution patterns of signals. In this article, a frame singular value decomposition (F-SVD) based method-generalizing Canonical correlation analysis for automatic classification of EMG signals to diagnose amyotrophic lateral sclerosis (ALS), myopathy and normal subjects, is proposed. At first, signals are decomposed to formulate a set of vectors and performed subspace transformation to demonstrate the variability and stability of signals base on correlations between pairs of vectors. Besides, discrete Wavelet transformation is applied on generated vectors and correlation analysis is performed. Afterwards, taking highly correlated statistical measures a set of compact feature distributions are estimated and fused via two recently proposed parallel and serial feature fusion models. Finally two global descriptors for effective classifications of various EMG patterns are proposed. The efficacy of derived feature space is validated by intuitive, graphical and statistical analysis. The model performances are investigated over two datasets. It achieves accuracy of 98.10% and 97.60% over two and three-class groups of first dataset receptively. Accuracy over second dataset is 100% with a specificity of 100% and sensitivities of 100%. This is first time that F-SVD is employed for automatic classification of EMG. Experiments results on various datasets evince adequacy of our method. Further comparison of performance with state-of-the-art methods depicts that our method comparable or superior in terms of various performance metrics.

Nonlinear Offset Measurement and Nullification for Effective Resistive Sensor Design

Over the decades, a number of methodologies have been introduced to meliorate the resistive sensor measurement protocol for complete knowledge of the phenomenon of interest. Nonetheless, such setting requires high degree of circuit components that result high level of errors (i.e., nonlinear) and thereby, its minimization for effective design is an open question. This article presents a technique that utilizes direct resistive circuit with microcontroller (\( \mu C \)), followed by subsequent estimation of curve-fitting models (CFMs) to curtail the errors involved and implementation in \( \mu C \) to update real-time data. Further, the study exploited the effectiveness of various employed CFMs in this context. The significant aftermaths with suitable choice of CFM and subsequent comparison with the state-of-the-art approaches manifest the efficacy of the adopted scheme.

An automatic feature extraction and fusion model: application to electromyogram (EMG) signal classification

In this paper, we present a real-time feature extraction and fusion model for an automated staging of electromyogram signals-generalizing canonical correlation analysis (CCA). The proposed method is capable of capturing multiple view information (i.e., feature matrices) generated from signals. Our algorithm employs an optimization technique to derive sets of statistical features among the paired views based on which possible variations of signals have been demonstrated. Next, discrete wavelet transformation is performed on multiple views to create domain independent views which are then subjected to CCA optimization. The estimated two sets of statistically independent features from two independent analysis are concentrated through two recently proposed fusion models, and then, we evaluate global feature matrices. Further it is validated statistically for

Effect of Annealing Temperature on the Morphology and Sensitivity of the Zinc Oxide Nanorods-Based Methane Senor

p<0.05

Nonlinearity compensation of DIC-based multi-sensor measurement

p<0.05. The proposed algorithm is then analyzed and compared with state-of-the-art methods. Results indicate that the proposed approach outperforms many other methods in terms of accuracy, specificity and sensitivity, which are 98.80, 99.0 and 98.0%, respectively. Thus, the proposed algorithm is suitable for large-scale applications and expedite diagnosis research.