Rocio Alba-Flores

Detection and Classification of Cardiac Murmurs using Segmentation Techniques and Artificial Neural Networks

In this paper we present the implementation of a diagnostic system based on artificial neural networks (ANN) that can be used in the detection and classification of heart murmurs. Segmentation and alignment algorithms serve as important pre-processing steps before heart sounds are applied to the ANN structure. The system enables users to create a classifier that can be trained to detect virtually any desired target set of heart sounds. The output of the system is the classification of the sound as either normal or a type of heart murmur. The ultimate goal of this research is to implement a heart sounds diagnostic system that can be used to help physicians in the auscultation of patients and to reduce the number of unnecessary echocardiograms - those that are ordered for healthy patients. Testing has been conducted using both simulated and recorded patient heart sounds as input. Three sets of results for the tested system are included herein, corresponding to three different target sets of simulated heart sounds. The system is able to classify with up to 85 plusmn 7.4% accuracy and 95 plusmn 6.8% sensitivity. For each target set, the accuracy rate of the ANN system is compared to the accuracy rate of a group of 2nd year medical students who were asked to classify heart sounds from the same group of heart sounds classified by the ANN system. System test results are also explored using recorded patient heart sounds

Recognition and classification of cardiac murmurs using ANN and segmentation

A diagnostic system based on Artificial Neural Networks (ANN) is implemented as a detector and classifier of heart murmurs. Segmentation and alignment algorithms serve as important pre-processing steps before heart sounds are applied to the ANN structure. The system enables users to create a classifier that can be trained to detect virtually any desired target set of heart sounds. The output of the system is the classification of the sound as either normal or a type of heart murmur. The ultimate goal of this research is to develop a tool that can be used to help physicians in the auscultation of patients and thereby reduce the number of unnecessary echocardiograms- those that are ordered for healthy patients. Testing has been conducted using both simulated and recorded patient heart sounds. Results are described for a system designed to classify heart sounds as normal, aortic stenosis, or aortic regurgitation. The system is able to classify with up to 85 ± 7.4% accuracy and 95 ± 6.8% sensitivity the tested heart sounds. Results are also described for a system designed to classify heart sounds using the consensus result of two sub-systems: (1) normal or aortic stenosis and (2) normal or aortic regurgitation. The consensus system is able to classify the same set of sounds with up to 96.8 ± 2.2% accuracy and 95.9 ± 5.2% sensitivity.

A case study on tuning artificial neural networks to recognize signal patterns of hand motions

This paper presents the development of artificial neural networks (ANN) as pattern recognition systems to classify surface electromyography signals (sEMG) into nine select hand motions from seven subjects. Multiple networks were designed to determine how well a network could adapt to signals from different subjects. This was achieved by developing multiple networks with different combinations of the volunteers for training. Each network was tested with signals from all volunteers to determine how well they could adapt to new subjects. It was found that ANNs trained using only one or two subjects would perform exceptionally well when tested with signals from the same subjects but relatively poorly when tested with signals from new subjects. As the number of subjects used for training increased, the ability of the network to accurately classify the signals from the trainees decreased but their ability to adapt to signals from new subjects increased. Solely based on these results, it can be inferred that ANNs developed using signals from a large amount of subjects could be used to accurately classify signals from completely new subjects. Research presented in this paper has potential to be further developed as a basis for utilizing sEMG as control signals in electric devices such as myoelectric prosthesis or humanoid control.

EMG Based Classification of Percentage of Maximum Voluntary Contraction Using Artificial Neural Networks

This paper presents an application of an Artificial Neural Network (ANN) for the classification of Electromyography (EMG) signals. The classification system has been designed to classify the percentage of maximum voluntary contraction (%MVC) from the bicep muscle. The EMG signals used in this study have been generated using a computer muscle model. Three statistical input features are extracted from the EMG signals and different structures of ANNs and training algorithms have been considered in the study. A 16 neuron hidden layer architecture trained with the scaled conjugate gradient algorithm has been found to be more efficient than the other ANN architectures tested in classifying 9 different bicep muscle contraction levels as a unit of %MVC than other ANN architectures. The ultimate goal of this research is to design a robotic system for people with disabilities and the elderly by utilizing muscle contraction levels as the input of tasks for the robot.

Performance analysis of two ANN based classifiers for EMG signals to identify hand motions

The aim of this work is to develop an accurate method for pattern recognition of human hand motions. Eight surface EMG electrodes (dual type) were placed on the forearm of healthy subjects while performing individual wrist and finger motions. A total of 1080 signals that incorporated all the selected nine hand motions were acquired from 12 volunteers, preprocessed, and then time-domain features were extracted. Two ANN architectures were developed and their performance was compared. The first architecture used a single ANN to perform the classification of the nine hand movements. This architecture achieved an average accuracy of all classes of 83.43%. In an effort to improve the accuracy of the classification, a second ANN architecture was developed. The second architecture consisted of nine independent ANNs, each one designed and trained to detect a specific hand motion. The second architecture achieved an average accuracy of all classes of 91.85%. Although the second ANN architecture showed an improvement in the accuracy, more research has to be performed before this type of ANN architectures can be used in real-life applications.

Qualitative evaluation of a PID controller for autonomous mobile robot navigation implemented in an FPGA card

This paper presents a new quantitative and qualitative approach for the design and implementation of a PID controller for autonomous mobile robot navigation. The paper also presents a simple and inexpensive method of designing, implementing, and evaluating the performance of the PID controller. The PID controller was first designed and simulated in Simulink. Then a FPGA card was used to build the PID controller in combination with a Basic Stamp microcontroller that is mainly used to perform I/O processing, sensor fusion, and data acquisition tasks. Finally, the performance of the PID controller, used to control the robots navigation, was evaluated in real environment situations using a rubric. Results demonstrated that the proposed approach provides a simplified and inexpensive method to design, implement, and evaluate the performance of controllers for mobile robot applications.

Relating movement primitive length to accuracy in imitation learning

We present an analysis of the effect of movement primitive length on accuracy using a visual externally observed imitation learning algorithm. Utilizing the noise reduction techniques in the chosen algorithm, we show that, though the largest impact on output path accuracy is the quality of the input data, the inaccuracy of the output increases as the movement primitive length increases as shown by the smoothing splines sum of squares goodness of fit and the number of data points rejected by the low pass filter.

Module detection for bacteria based on spectral clustering of protein-protein functional association networks

Network analysis-based module detection has significant implications in many fields. In cellular/ molecular biology, module detection based on analyses of metabolic/regulatory networks will not only help us understand more about the function and evolution of cellular machinery of an organism, but will also provide tractable contextual information for potential drug targets and facilitate improvements in drug designs. We here present our preliminary study on the module detection for bacteria based on the spectral clustering of the protein-protein functional association networks. We first examined how the parameter of the spectral clustering algorithm (i.e., the number of clusters) affects our module detection results, and demonstrated that when the number of clusters was set too small or too large the resulting module collection deteriorate in terms of gene coverage and intra-module association. We then compared our predicted modules against the randomly generated modules, and demonstrated that our modules (i) have a higher ratio of the intra-module to inter-module gene-gene functional association scores and (ii) can better capture the modularization information inherent in the experimentally verified modules. Finally we compared the module collections of seven bacterial organisms, and observed that modules related to membrane transport and cell motility are among those that are conserved among multiple organisms. Because it is desirable from both scientific and technical points of view to study functional modules at various resolution levels, we believe that the spectral clustering algorithm, with the flexibility rendered by different parameter settings, provides an appropriate solution in terms of capturing the modularization properties of networks and computational affordability.