# Detection and classification of respiration disorder based on breathing pattern using fuzzy min-max classifier.

INTRODUCTIONThe human respiratory system consists of list of organs and it is responsible for inhale and exhale oxygen and carbon dioxide respectively. The most important organs in respiratory system are lungs that carry out the exchange of gases as we breathe.

Respiratory disease includes medicinal conditions affecting the tissues and organs. It leads to upper respiratory tract, pleura, bronchi and the nerves and muscles of breathing. Respiratory illnesses run from gentle and self-restricting such as the common cold, life-threatening entities like bacterial pneumonia, pulmonary embolism, and lung cancer.

The main function of the respiratory system is breathing that supplies oxygen to all parts of the body. The respiratory system lies inactive in the human fetus during pregnancy. At birth, fully functional upon exposure to air, although the development of lung continues throughout childhood. In some cases pre-term birth can lead to infants with under-developed lungs. Mostly respiratory problems causes due to the smoking and air pollution. In this paper, different machine learning techniques are employed to classify the respiratory disorder.

Literature Survey:

Atena Roshan Fekr et al [1] proposed a wearable MEMS sensor technology. They detect eight types of breathing problem with one and two accelerometer accordingly. The detected data streams are divided into fixed size overlapping and non-overlapping window which are represented as class. The features are extracted by using correlation feature selection filter method. However the effective breathing therapy is not discussed.

F. Mancini et al [2] proposed LVQ classification method to determine features from postural profiles that are taken by mouth breathing children. This LVQ classification method is used only for postural profile shown by mouth breathing children.

Majid Janidarmian et al [3] proposed Analysis of Motion Patterns for Recognition of Human Activities. Accelerometer sensors are sensed data by mounting it on human body. However there is no feature selection process in order to select specific features.

Julien Oster et al [4] proposed Semi-supervised ECG Ventricular Beat Classification with Novelty Detection Based on Switching Kalman Filters. A Switching Kalman Filter technique is used to enable the automatic selection of mostly like heartbeats while simultaneously filtering the signal with prior knowledge. Switching Kalman Filter is more complicated in noisy environments.

YEE SIONG LEE et al [5] proposed Monitoring and Analysis of Respiratory Patterns Using Microwave Doppler Radar. This technique avoids the use of physical sensors that are mounted on patient's body. Microwave Doppler radar is used to capture various dynamics of breathing patterns including respiration rate. But Doppler radar may be affected by environmental conditions.

Peter Varady et al [6] proposed an innovative signal classification method that is capable of online detection of presence or absence of normal breathing. Here four different neural networks are presented to recognize the different patterns in respiration signals. It detects signal without using patient specific information and moderates computational time. But heart rate and large signal database are needed.

Mila Kwiatkowska et al proposed [7] analysis of informational and technological requirements for the respiratory therapy workshops in Peru. This paper investigates the use of low cost and low resource methods for at-location and learner-centered medical education and training. This paper helps to us to learn about medical therapy for avoiding respiratory disorders.

SASAN AHDI REZAEIEH et al [8] proposed the design and implementation of an automated ultrahigh frequency microwave based system for CHF detection and monitoring is presented. The collected data from the scanning is then visualized in the time domain using the inverse Fourier transform. Using a differential based detection technique a threshold is defined to differentiate between healthy and unhealthy cases.

Ching-Wei Wang et al [9] proposed unconstrained video monitoring of breathing behaviour and application to diagnosis of sleep Apnea. The algorithm uses a novel persistence luminance model that helps to reinforce subtle breathing movements, an activity level to segment the video, and a novel activity template to classify motion events. Human breath activity by adding audio and human sleep behaviour are not discussed in this technique.

Yingying Zhenga et al [10] proposed predicting arterial stiffness from radial pulse waveform using support vector machines. Here radial pulse signals are decomposed into time-frequency representations using DWT and wavelet scale-energy are calculated. This is used to predict optimum classification method. This features achieved high classification accuracies.

Proposed Methodology:

LVQ Classification Method sensors are used to detect eight types of breathing problems with modelling the small movements of the chest wall compartments that occur during expansion and contraction of the lungs in each respiration cycle. Initially sensors are mounted on patient's rib cage and abdomen and this sensors detects various types of breath disorders such as Bradypnea, Tachypnea, Kussmaul, Cheyn-stokes, Obstructive Sleep Apnea (OSA), Biot's breathing, Sighing and Apneustic. Sensed data streams are dividing into Fixed-size Nonoverlapping Window (FNSW) and Fixed-size Overlapping Window (FOSW). Then the features such as Mean, Standard Deviation, Respiration rate and time, Total variability, Number of peaks are extracted. The features are selected by using filter method called Correlation-based Feature Selection (CFS). The selected features are classified by using various algorithms including Decision Tree Bagging (DTB), k-Nearest Neighbors (KNN), Support Vector Machines (SVM) and Artificial Neural Networks (ANN). In this proposed system we include two more classification algorithms namely Learning vector Quantization (LVQ) and Modified Fuzzy Min-Max Classifier using compensatory Neurons (M-FMCN).These features are transferred to cloud through interfaces by using BLE. Then physicians can access the data which are stored in cloud. By using this technique Breathing Therapy is an effective and drug-free treatment technique. This provides strong immune system, better bloodsugar control and etc. Cost of diagnosis and medical care are less. LVQ classification method has high accuracy, sensitivity and specificity.

A. Data Segmentations:

For data segmentation, two window-based segmentation methods are used with respect to the window length and the percentage of adjacent windows overlap in obtaining the highest classification accuracy. The first method is dividing the data stream into windows of fixed length with no inter-window gaps and no degree of overlap between adjacent windows (Fixed-size Non-overlapping Sliding Window (FNSW)). The second method is Fixed size Overlapping Sliding Window (FOSW) in which windows overlap during segmentation. Then the features are extracted from each separate window of data and then used as input to the classifiers. The selected features include Mean, Standard Deviation (SD), and respiration rate, respiratory time parameters, tilt angles, tidal volume variability, accelerometer-based breathe volume, phase shift, and symbolic aggregate approximation (SAX) of the data.

B. Feature extraction:

In order to get high accuracy for sensor measurements, calibration technique is performed using least square method. To calibrate an accelerometer is performed at 6 stationary positions. A few seconds of accelerometer raw data at each position is collected. Then the least square method is applied to obtain 12 accelerometer calibration parameters. Then the sensory data requires segmentation in order to facilitate effective feature extraction technique.

C. Feature Selection:

The features are selected by using Correlation-based Feature Selection (CFS) which is categorized as a filter technique. The following equation gives the heuristic merit of a feature subset S consisting of k features:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (1)

In above equation, [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] is the average feature-class correlation and [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] the average feature-feature correlation. CFS can start from an empty set of features and uses a forward best first search.

C. Classification Algorithms:

In our proposed system, two classification techniques are proposed in order to get high accuracy and improve the cloud storage for signal processing. Thus, after data segmentation and feature selection, the training set is labeled, determining the corresponding classes.

D. LVQ classification:

In this classification, initially assign weight vectors to the first m training vectors, where m is the number of different categories and set a(0). For each training input vector find the new vector J in which Euclidean distance is a minimum and updates their weights as follows: Input vector, F = ([f.sub.l], [f.sub.2], [f.sub.3], ..., [f.sub.n]) (2)

Weight vector for jth vector is,

[w.sub.j] = ([w.sub.1j], [w.sub.2j], ..., [w.sub.nj]) (3)

Euclidean distance between the input vector and weight vector is denoted as:

D(j) = [square root of [[sigma].sup.n.sub.i=1][([P.sub.i] - [W.sub.ij]).sup.2] (4)

The correct category for input F is T and Category represented by jth vector is [C.sub.j]. If T [not equal to] [C.sub.j] then update the weights of J vector as:

[W.sub.j](new) = [W.sub.j](old) + a(F - [W.sub.j](old)) (5)

The above equation defined as move the weight vector toward the input vector F. If T [not equal to] [C.sub.j] then update weights of J vector as:

[W.sub.j](new) = [W.sub.j](old) - a(F - [W.sub.j](old)) (6)

The above equation defined as move w away from F. Then reduce the learning rate a. The classification may be a fixed number of iterations or the learning rate reaching a sufficiently small value.

Algorithm:

1. Initialize input vector,,F = ([f.sub.1],[f.sub.2],[f.sub.3], ..., [f.sub.n])

2. Compute weight vector for jth vector, [W.sub.j] = ([w.sub.1j], [w.sub.2j], ..., [w.sub.nj])

3. [C.sub.j] = Category represented by jth vector

4. T = Correct category for input F

5. Define Euclidean distance between the input vector and weight vector as:

d(j)= [square root of [n.summation over (i=1)][([P.sub.i] - [w.sub.ij]).sup.2]

E. Fuzzy Min-Max Classification:

In this classification, the feature vector extracted from the input is considered as input vector to the neural network. This input vector is applied to the input layer for the neural network and number of nodes in the input layer is equal to the dimension of applied input vector Ah. Where the [A.sub.h1], [A.sub.h2], ..., [A.sub.hn] are the input sample belongs to the pattern area [I.sub.n]. And [A.sub.1],[A.sub.2], ..., [A.sub.n] are the corresponding input nodes. The second layer neuron called hyper-box nodes [B.sub.1],[B.sub.2], ..., [sub.Bj] are created at the training time, which represents the Min-Max points of the hyper-box and are stored into the (V, W) matrix. The activation function [B.sub.j] is given as:

[B.sub.j] = {[A.sub.h],[V.sub.j], [W.sub.jf]([A.sub.h],[V.sub.j],[W.sub.j])} where [A.sub.h] [member of] [I.sup.n]

> Min point of jth Hyper-box is, [V.sub.j] = ([V.sub.j1],[V.sub.j2], ..., [V.sub.jn])

> Max point of jth Hyper-box is, [W.sub.j] = ([W.sub.j1], [W.sub.j2], ..., [W.sub.jn])

> Input vector of jth Hyper-box is, [W.sub.j] = ([A.sub.h1], [A.sub.h2], ..., [A.sub.hn])

> Number of dimension is n

The membership function for the jth Hyper-box is, 0 [less than or equal to] [B.sub.j]([A.sub.h], [V.sub.j],[W.sub.j]) [less than or equal to] 1

In this classifier, we assume that the degree of membership of input vector [A.sub.h] for the hyper-box is [B.sub.j] one if [A.sub.h] is in or within the hyper-box [B.sub.j] and the degree of membership decreases as [A.sub.h] moves away from the hyperbox [B.sub.j]. Thus, the new activation function is given as:

[b.sub.j] = ([A.sub.h] [v.sub.j] [w.sub.j]) = min(min [(l - f([a.sub.hi] - [w.sub.ji], [gamma])), ((l - f{[v.sub.ji] - hi, [gamma]) (7)

for i = 1, ..., n (7)

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (8)

Where [gamma] = ([[gamma].sub.1],[[gamma].sub.2], ..., [y.sub.n]) are sensitivity parameters regulating how fast the membership values decrease. This activation function is assigned if a input sample is within the hyper-box. Otherwise its membership function calculation is based on the distances between Min-Max points of the pattern A. Hyper-box nodes in Main Section are created if training sample belongs to a class. This has not been encountered that the existing hyper-boxes of that class cannot be expanded further to accommodate it. The connections between hyper-box node and class node in main section are represented by matrix U and hence the learning process is improved. The OCN takes over the control from the overlapped hyper-boxes and assigns membership to the test sample depending upon its distance from the min-max points.

The hyper-box nodes in middle layer OCN section are created whenever the network faces problem of overlap or containment. The OCN section takes care of the overlap problem. The connections between hyperbox and class nodes in OCN section are represented by matrix Y. The connection weight from neuron [d.sub.p] is representing the overlap between ith and jth class hyperbox. A connection between hyper-box nodes to a class node is adjusted by the following equation.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (9)

OCN Produces two yields, one each for the two covering classes. OCN is dynamic just when a test has a place with the cover distinct. The activation function is given as:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (10)

The Activation function of this neuron is such that it protects the class of the min and max point of overlapped hyper-boxes. The class node in OCN section is given by,

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (11)

Thus the classification is done by modified FMCN technique.

Algorithm:

1. Initialize the input vector and input samples.

2. Create hyper-box nodes and Min-Max points.

3. Compute activation function [B.sub.j].

4. Determine the new activation function [b.sub.j].

5. Adjust the connection between the hyper-box nodes.

6. Find activation function for overlapping hyper-box classes.

Thus our proposed system classifies the various respiration disorders accurately.

RESULTS AND DISCUSSION

This section presents the experimental results that are performed to prove the proposed fuzzy min-max classification technique achieves high accuracy. The performance of the proposed feature classification is evaluated in terms of precision, recall, accuracy and f-measure with existing classification techniques such as KNN, ANN, SVM and DTB classifiers. The performance of proposed classification techniques are simulated by using MATLAB.

A. Precision:

Precision value is evaluated according to the feature classification at true positive prediction; false positive. It is expressed as follows:

Precision = True positive/True positive + False positive

Fig. 1 shows that the comparison of the existing classification methods such as KNN, ANN, SVM and DTB with proposed methods LVQ and Fuzzy min-max classifications in terms of precision values. The result shows that the fuzzy min-max classification method provides higher precision than other classification techniques.

B. Recall:

Recall is assessed by the element characterization at genuine positive expectation, false negative. It is given as,

Recall = Truepositive/(Truepositive + Falsenegative)

Fig. 2 shows that the comparison of the existing classification methods such as KNN, ANN, SVM and DTB with proposed methods LVQ and Fuzzy min-max classifications in terms of recall values. The result shows that the fuzzy min-max classification method provides high recall than other classification techniques.

C. F-Measure:

F-measure is calculated from the precision and recall value. It is calculated as: precision x recall

f - measure = 2 x (precision x recall/(precision + recall))

Fig. 3 show that the comparison of the existing classification methods such as KNN, ANN, SVM and DTB with proposed methods LVQ and Fuzzy min-max classifications in terms of f-measure. The result shows that the fuzzy min-max classification method provides higher f-measure than other classification techniques.

D. Accuracy:

The accuracy is the proportion of true results (both true positives and true negatives) among the total number of cases examined. Accuracy can be calculated from formula given as follows

Accuracy = True positive + True negative/True positive + True negative + False positive + False negative

Fig. 4 show that the comparison of the existing classification methods such as KNN, ANN, SVM and DTB with proposed methods LVQ and Fuzzy min-max classifications in terms of accuracy. The result shows that the fuzzy min-max classification method provides higher accuracy than other classification techniques.

E. Time Taken:

Fig. 5 show that the comparison of the existing classification methods such as KNN, ANN, SVM and DTB with proposed methods LVQ and Fuzzy min-max classifications in terms of time. The result shows that the fuzzy min-max classification method provides lesser time than other classification technique.

Conclusion And Future Work:

This paper presented an efficient classification method for classifying the different types of respiration disorder patterns. The features are classified based on two different classification techniques such as LVQ and Fuzzy min-max technique to improve the classification accuracy and sensitivity. The experimental results are demonstrated that the fuzzy min-max classifier has better accuracy and sensitivity than other classification techniques. Comparison results illustrates that our proposed system outperforms than the existing system. The drug-free breathing therapy such as breathing methods helps patients to recover from their respiratory disorders without any effect.

The future extensions of this research are to transmit the sensory data securely by introducing security mechanisms. Also, IoT will be added for effective real time applications. Moreover, this technique will be extended towards big data analytics to analyse the large set of respiration patterns.

REFERENCES

[1.] Fekr, A.R., M. Janidarmian, K. Radecka and Z. Zilic, 2016. Respiration Disorders Classification with Informative Features for m-health Applications. IEEE journal of biomedical and health informatics, 20(3): 733-747.

[2.] Mancini, F., F.S. Sousa, A.D. Hummel, A.E.J. Falcao, L.C. Yi, C.F. Ortolani and I.T. Pisa, 2011. Classification of pSostural profiles among mouth-breathing children by learning vector quantization. Methods of information in medicine, 50(4): 349.

[3.] Janidarmian, M., A. Roshan Fekr, K. Radecka, Z. Zilic and L. Ross, 2015. Analysis of Motion Patterns for Recognition of Human Activities. In Proceedings of the 5th EAI International Conference on Wireless Mobile Communication and Healthcare (pp. 68-72). ICST (Institute for Computer Sciences, SocialInformatics and Telecommunications Engineering).

[4.] Oster, J., J. Behar, O. Sayadi, S. Nemati, A.E. Johnson and G.D. Clifford, 2015. Semisupervised ECG ventricular beat classification with novelty detection based on switching Kalman filters. IEEE Transactions on Biomedical Engineering, 62(9): 2125-2134.

[5.] Lee, Y.S., P.N.P athirana, C.L. Steinfort and T. Caelli, 2014. Monitoring and analysis of respiratory patterns using microwave doppler radar. IEEE journal of translational engineering in health and medicine, 2: 1-12.

[6.] Varady, P., T. Micsik, S. Benedek and Z. Benyo, 2002. A novel method for the detection of apnea and hypopnea events in respiration signals. IEEE Transactions on Biomedical Engineering, 49(9): 936-942.

[7.] Kwiatkowska, M. and L.M atthews, 2014. Analysis of informational and technological requirements for the respiratory therapy workshops in Peru. In Global Humanitarian Technology Conference (GHTC), 2014 IEEE pp: 675-681.

[8.] Rezaeieh, S.A., K.S. Bialkowski and A.M. Abbosh, 2014. Microwave system for the early stage detection of congestive heart failure. IEEE Access, 2: 921-929.

[9.] Wang, C.W., A. Hunter, N. Gravill and S. Matusiewicz, 2014. Unconstrained video monitoring of breathing behavior and application to diagnosis of sleep apnea. IEEE Transactions on Biomedical Engineering, 61(2), 396-404.

[10.] Zheng, Y., Y. Zhang, Z. Ma and Y. Sun, 2010. Predicting Arterial Stiffness from radial Pulse Waveform using support vector machines. Procedia Engineering, 7: 458-462.

(1) V. Shanmathi and (2) Dr. S. Jeyanthi

(1) PG Scholar, Department of Computer Science and Engineering, PSNA College of Engineering and Technology, Dindigul

(2) Assistant Professor, Department of Computer Science and Engineering, PSNA College of Engineering and Technoiogy, Dindigui.

Received 28 January 2017; Accepted 22 May 2017; Available online 28 May 2017

Address For Correspondence:

V. Shanmathi, PG Scholar, Department of Computer Science and Engineering, PSNA College of Engineering and Technology, Dindigul.

E-mail: shanmathiv54@gmail.com

Caption: Fig. 1: Comparison of precision

Caption: Fig. 2: Comparison of recall

Caption: Fig. 3: Comparison of f-measure

Caption: Fig. 4: Comparison of accuracy

Caption: Fig. 5: Comparison of time taken

Printer friendly Cite/link Email Feedback | |

Author: | Shanmathi, V.; Jeyanthi, S. |
---|---|

Publication: | Advances in Natural and Applied Sciences |

Article Type: | Report |

Geographic Code: | 9INDI |

Date: | May 1, 2017 |

Words: | 3326 |

Previous Article: | A modified Frank Wolfe algorithm for social recommendation based on user preference in multimedia application. |

Next Article: | A fuzzy rule based decision making algorithm for detection of flood event. |

Topics: |