4

Click here to load reader

Fully Connected CNN for ECG Recognition on DSP Connected CNN... · Fully Connected CNN for ECG Recognition on DSP Javeed Ahmed Khan1, S. Ravichandran2, ... ˝Digital Signal Processing-A

  • Upload
    lytu

  • View
    216

  • Download
    4

Embed Size (px)

Citation preview

Page 1: Fully Connected CNN for ECG Recognition on DSP Connected CNN... · Fully Connected CNN for ECG Recognition on DSP Javeed Ahmed Khan1, S. Ravichandran2, ... ˝Digital Signal Processing-A

TECHNIA – International Journal of Computing Science and Communication Technologies, VOL. 2, NO. 1, July 2009. (ISSN 0974-3375)

273

Fully Connected CNN for ECG Recognition onDSP

Javeed Ahmed Khan1, S. Ravichandran2, and Gopalakrishna.K3

1Dr. MGR University, Chennai.2Trimentus Technologies, Chennai – 17.

3DIAS, Bangalore.([email protected])

Abstract:- This paper describes about the work carriedout in the field of telemedicine using InformationTechnology (IT), which is the use of dedicated ITprocessors and telecommunications for medicaldiagnosis and patient care. For constant patientmonitoring and diagnosis this method is devised,however for full scale practical use it requires furtherenhancing tests. Telecommunication is carried out byInternet Protocol (IP) which transmits the ElectroCardio Gram (ECG) through a device at the patientside. At the receiving end, the signal isreceived/acquired on to the physician’s computerthrough for analysis. The software is Cellular NeuralNetwork implemented on Digital Signal Processor.Object Recognition is performed to check theauthenticity of the received signal, higher no. oftraining sets reduces the time for convergence andhence can be used for authentication with fasterconvergence and least human intervention.

Keywords: Electro Cardio Gram (ECG), ObjectRecognition (OR), Internet Protocol (IP), Neural Network(NN), Cellular Neural Networks (CNN), Multiply andAccumulate (MAC) and Digital Signal Processor (DSP).

1. INTRODUCTION

ECG is acquired at the patients end using three leadsconnected as per standard methodology followed by allpractitioners; signals received will be then compressed andtransmitted through Internet Protocol from the patient tothe doctor/physician. The software/algorithmimplemented at this end will be having more ofindependent approach towards the analysis as it will haveless human intervention. Because of the fact that thephysician’s computer houses more no. of such data andconstant monitoring is required. This method givesconstant monitoring using Object Recognition techniques,and with Cellular Neural Networks the interpretation taskis minimized as it learns and acquires knowledge on itsown after adequate training. Digital Signal Processor likeTexas Instruments (TMS320C542) is preferred forprocessing work on which the CNN Program is loaded forthe current work carried out at one real physicians place.

Object recognition paradigms are repetitive Additionand Multiplication algorithms for which dedicated

processors with Super Harvard Architecture(SHARK)/Multiply and Accumulate (MAC) architectureis used. Advancements in computer technology along withartificial intelligence paradigms developed have madehuman intervention limited and matching of ECG’s isaccomplished with these algorithms working on dedicatedhardwares which reduce the time for processing.

Fig 1.1 shows the block diagram for monitoring.

Leads from patient

Fig 1.1 Block diagram of Experimental setup

Power Supply Block (12V, 500mA)

High I/pImpd.

Differential

DifferentialAmplifier

BufferAmplifier

IPPatient

IP Physician

Processing&

Interpreting Unit

Page 2: Fully Connected CNN for ECG Recognition on DSP Connected CNN... · Fully Connected CNN for ECG Recognition on DSP Javeed Ahmed Khan1, S. Ravichandran2, ... ˝Digital Signal Processing-A

TECHNIA – International Journal of Computing Science and Communication Technologies, VOL. 2, NO. 1, July 2009. (ISSN 0974-3375)

274

V V

Vr

g g gr

icir

is

R

Cm

is

2. CELLULAR NEURAL NETWORKS

Cellular Neural Networks is an information-processingparadigm that is inspired by the way biological nervoussystems, such as the brain processes information. Fig 2.1shows an electrical equivalent of a neuron/single cell.

Fig 2.1 Electrical Equivalent of Neuron/Single Cell

In here the classification of ECG data is done into fourtwo groups, viz. Normal ECG and Abnormal ECG.

Mathematical modeling of the analysis toolabout CNN is shown. Equations for CNN are indifferential equation form.

Standard CNN differential equation is shown ineqn. 2.1

r

rk

r

rk

r

rlljkiklljki

r

rlkl zubyaXijXij ,,

.

……… eqn (2.1)

Where.

Xij is the first derivative of Xij . ‘a’ and ‘b’ are the elements of the space invarianttemplate matrices.

Solving of standard differential equation like.x =h(x;w)

x=x(t)

x(0)=x0 ………. eqn (2.2) This can be solved by standard numerical integrationmethods; the simples one is the forward Euler formulawhich calculates the value of x(t+ )t from x(t), tbeing the time step.

x(t+ )t ));(()()()( wtxthtxtxttx ……… eqn (2.3)

From, eqn(2.3) it is qualitatively correct and accurateenough if we use a small time step t , small enough thatthe CNN dynamics range when known in advance stateswill give good convergence, advance states meaning thegroup to which the PD pattern belongs.

The above equations 2.1 to 2.3 can be represented infigure 2.2 form as shown below.

Fig 2.2 A two dimensional fully connected Cellular Neural Network

In figure 2.2, it is seen that the basic circuit unitof Cellular Neural Networks is called a cell. It containslinear and nonlinear circuit elements. The structure ofCellular Neural Networks is similar to that found inCellular automata: namely, any cell in a cellular neuralnetwork is connected only to its neighbour cells. Theconnected cells can interact directly with each other [1].Cells not directly connected together may affect eachother indirectly because of the propagation effects of thecontinuous time dynamics of cellular neural networks.

Consider an MxN Cellular Neural Network,having MxN cells arranged in M rows and N columns.We call the cell on the ith and the jth column cell (i,j), anddenote it C(i,j) as in figure 2.2.

3. DIGITAL SIGNAL PROCESSOR

Generic Processors use Von-neumannarchitecture, which uses single data bus and address busand is multiplexed this hampers the processing speed forrepetitive processes like the one represented in eqn. 2.1 [4][5][8][9]. Digital Signal Processors are made on SuperHarvard Architecture (SHARK) which has data bus andaddress bus separate and that the stored program andstored data are separate but interlinked. This is ideal forrepetitive task like the multiply and accumulate task

C (1,1) C (1,2) C (1,3) C (1,4)

C (2,1) C (2,2) C (2,3) C (2,4)

C (3,1) C (3,2) C (3,3) C (3,4)

C (4,1) C (4,2) C (4,3) C (4,4)

Page 3: Fully Connected CNN for ECG Recognition on DSP Connected CNN... · Fully Connected CNN for ECG Recognition on DSP Javeed Ahmed Khan1, S. Ravichandran2, ... ˝Digital Signal Processing-A

Khan et al.: Fully Connected CNN for ECG Recognition on DSP

275

(MAC) [2][7]. The DSP used here for processing, i.e.,computing the distances or weights after repetitive feedforward training was of fixed point origin, the softwaredeveloped for DSP was written in high level language andlater converted to assembly level language and finally tomachine level language.

4. METHODOLOGY

Fig 1.1 shows the block diagram representation of theMonitoring system. However, the methodology forprocessing the data is discussed as follows. The ECGsignal obtained from the patient is as shown in Fig 4.1 isconverted into its pixel value and stored in the matrix form[3].

Fig 4.1 Normal patient ECG

Fig 4.2 to 4.4 shows the process of converting theimage into its equivalent pixel values.

Fig 4.2 Image of ECG signal in pixel value

Fig 4.3 Image of ECG signal in its original form

Fig 4.4 Image of ECG signal in its final pixel value

Fig 4.4 shows the pixel value which is stored in thematrix form and transmitted over the IP and at the receiverend is compared with the normal ECG signal image, ifthere are ambiguities, the same is report to the physician[6].

5. RESULTS AND DISCUSSIONS

Table 5.1 shows the results obtained after analysis on anormal patients ECG. After repetitive iterations the no. ofiterations required will be reduced. It is shown in Table 5.2 and5.3 respectively.

TABLE 5.1 RESULTS OF NORMAL PATIENTSECG

LearningRate ( )

Momentum( )

No. OfIterations

Output/Convergence

0.1 0.5 450 0.90.2 0.6 250 0.8999990.3 0.7 250 0.9009000.4 0.8 350 0.9009000.5 0.9 100 0.90022

Page 4: Fully Connected CNN for ECG Recognition on DSP Connected CNN... · Fully Connected CNN for ECG Recognition on DSP Javeed Ahmed Khan1, S. Ravichandran2, ... ˝Digital Signal Processing-A

TECHNIA – International Journal of Computing Science and Communication Technologies, VOL. 2, NO. 1, July 2009. (ISSN 0974-3375)

276

TABLE 5.2 RESULTS OF NORMAL PATIENTS ECG AFTERREPETATIVE ITERATIONS

LearningRate ( )

Momentum( )

No. OfIterations

Output/Convergence

0.1 0.5 130 0.90.2 0.6 130 0.90.3 0.7 130 0.90.4 0.8 100 0.90.5 0.9 90 0.9

TABLE 5.3 RESULTS OF PARAMETRIC ECG ANALYSESFOR NORMAL PATIENT

LearningRate ( )

Momentum( )

No. OfIterations

Output/Convergence

0.1 0.5 50 0.90.2 0.6 50 0.90.3 0.7 50 0.90.4 0.8 48 0.90.5 0.9 45 0.9

Conclusion drawn from the results tabulated in table 5.1 to5.3 show the faster and accurate convergence. Table 5.4 givesthe comparison of results between normal ECG signal and ECGobtained from patients with abnormalities.

TABLE 5.4 RESULTS OF COMPARISON BETWEENNORMAL ECG AND ECG WITH ABNORMALITIES

Normal ECG data Epochs ECG data withabnormalities

DesiredOutput

ActualOutput

Iterations DesiredOutput

ActualOutput

0.99 0.9782 500 0.9 0.14110.99 0.9771 500 0.9 0.16520.99 0.9656 500 0.9 0.14290.99 0.9786 500 0.9 0.13260.99 0.9092 500 0.9 0.21350.99 0.9092 500 0.9 0.12530.99 0.974766 500 0.9 0.150.99 0.974766 500 0.9 0.12547110.99 0.974766 500 0.9 0.330.99 0.976369 500 0.9 0.37150.99 0.976369 500 0.9 0.332750.99 0.9776369 500 0.9 0.2612

6. CONCLUSIONS

Conclusions drawn from the above results suggest thatECG can also be transmitted and analyzed as pictorialrepresentation and can provide better results in this form.Error content associated with the ECG signal is reduceddrastically at all different levels. Except at patients end duringthe acquisition stage. However, during transmission andreception stage this error is drastically reduced and hence betterand accurate convergence is obtained.

Trained CNN can distinguish between normal andabnormal ECG’s and provide results for corrective steps tobe initiated.

REFERENCES

[1] Chuo. O., Young.L., “Cellular Neural NetworkTheory”, International Journal of Computer Vision, pp.77-99, 1996.

[2] Ifeachor. E.C & Jervis. B. W, “Digital SignalProcessing: A Practical Approach”, Addison-Wesley,1993.

[3] Javeed Ahmed Khan, “ECG Transmission throughMobile Phones”, Guidance to Dissertation work ofM.Tech student to VTU, 2004.

[4] Emmanuel C. Ifeachor and Barrie W. Jervis, “DigitalSignal Processing-A Practical Approach, Second edition,Pearson Education publication.

[5] Golub.G.H and Van Loan. C. F. “MatrixComputations”, (3rd Edition). Johns Hopkins UniversityPress, Baltimore, MD.

[6] Azevedo.S & Longini.R.L, “Abdominal-Lead FetalElectrocardiographic R-wave Enhancement of Heart RateDetermination”. IEEE Transactions on BiomedicalEngineering, vol.27, no.5, pp.255-260.

[7] Kunihiko Fukushima, “A Self-organizing multilayeredneural network”, Journal of Biological Cybernetics, vol.20, pp121-136, March 1975.

[8] Kamil A. Grajski, Dan P. Witmer and Carson Chen, “Acombined DSP and artificial neural network (ANN)approach to the classification of time series data”exponent: Ford Aerospace Technical Journal, pp.20-25,Winter 1989/1990.

[9] Robert Hecht-Nielsen, “Nearest matched filterclassification of spatiotemporal patterns”. Technicalreport, Hecht-Nielsen Neurocomputer Corporation, SanDiego CA, June 1986.