Error Correction Algorithm for High Accuracy Bio-Impedance Measurement in Wearable Healthcare...

196 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 8, NO. 2, APRIL 2014

Error Correction Algorithm for High AccuracyBio-Impedance Measurement in Wearable

Healthcare ApplicationsRajkumar Kubendran, Member, IEEE, Seulki Lee, Student Member, IEEE, Srinjoy Mitra, Member, IEEE, and

Refet Firat Yazicioglu, Member, IEEE

Abstract—Implantable and ambulatory measurement of physi-ological signals such as Bio-impedance using miniature biomedicaldevices needs careful tradeoff between limited power budget,measurement accuracy and complexity of implementation. Thispaper addresses this tradeoff through an extensive analysis ofdifferent stimulation and demodulation techniques for accurateBio-impedance measurement. Three cases are considered forrigorous analysis of a generic impedance model, with multiplepoles, which is stimulated using a square/sinusoidal current anddemodulated using square/sinusoidal clock. For each case, theerror in determining pole parameters (resistance and capaci-tance) is derived and compared. An error correction algorithmis proposed for square wave demodulation which reduces thepeak estimation error from 9.3% to 1.3% for a simple tissuemodel. Simulation results in Matlab using ideal RC values showan average accuracy of for single pole andfor two pole RC networks. Measurements using ideal componentsfor a single pole model gives an overall andreadings from saline phantom solution (primarily resistive) givesan . A Figure of Merit is derived based onability to accurately resolve multiple poles in unknown impedancewith minimal measurement points per decade, for given frequencyrange and supply current budget. This analysis is used to arrive atan optimal tradeoff between accuracy and power. Results indicatethat the algorithm is generic and can be used for any applicationthat involves resolving poles of an unknown impedance. It can beimplemented as a post-processing technique for error correctionor even incorporated into wearable signal monitoring ICs.

Index Terms—Bio-impedance, biomedical devices, error correc-tion, figure of merit.

I. INTRODUCTION

B IO-IMPEDANCE is the impedance of the human tissuewhich can be used for medical diagnosis in physio-

logical signal monitoring applications. It has various clinical

Manuscript received September 19, 2013; revised December 09, 2013; ac-cepted March 02, 2014. Date of publication April 25, 2014; date of current ver-sion May 23, 2014. This work was done as a part of the International VisitingScholar Internship supported by imec, Belgium. This paper was recommendedby Associate Editor K.-T. Tang.R. Kubendran was with Purdue University, West Lafayette, IN 47906 USA.

He is now with Qualcomm Technologies Inc., Boxborough, MA 01752 USA(e-mail: [email protected]).S. Lee is with Holst Centre/imec, Netherlands, 5656 AE Eindhoven, The

Netherlands (e-mail: [email protected]).S. Mitra and R. F. Yazicioglu are with imec, 3001 Heverlee, Belgium (e-mail:

[email protected]; [email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TBCAS.2014.2310895

applications ranging from respiration monitoring and tho-racic impedance measurements to body fluid analysis and cellcharacterization [1]–[10]. Since the amount of variation ofBio-impedance is very small to detect, an accurate currentsource and readout channel are necessary for precise mea-surement. Conventionally, a sinusoidal current is generated tomeasure the impedance, but accurate sine current generationusing a high resolution DAC, consumes large power and hencenot suitable for portable or wearable healthcare systems. Inorder to reduce power consumption, a square wave currentcan be used instead of the sine wave current. But, since thesquare wave current consists of many frequency componentsrather than a single tone frequency, the measurement accuracywill be degraded due to harmonic frequencies demodulated toproduce dc error. In this paper, impedance measurement usingthree different current stimulation-demodulation techniques –sine-sine, sine-square and square-square are analyzed and anerror correction algorithm is proposed to improve accuracy forthe sine-square and the square-square cases. This method canthen be adopted for wearable healthcare applications to achievehigh accuracy Bio-impedance measurements in low-powerminiature devices.This paper is organized as follows. In Section II, the setup

for Bio-impedance measurement using a custom ASIC is de-scribed. In Section III, a mathematical analysis is used to derivethe expressions for quadrature demodulated output voltagesmeasured from the ASIC described in [4]. The derivation isrepeated for all three cases. In Section IV, a measurementalgorithm is presented for square-square case which can beused to estimate the tissue impedance model parameters (resis-tance and capacitance) from the measured voltages at differentfrequencies [5]. Section V elaborates the simulation resultsfrom Matlab to verify the equations. Then measured resultsfrom saline solutions and ideal RC components are shownto validate the square-square error correction algorithm interms of accuracy and robustness. In Section VI, the trade-offsinvolved in considering better measurement accuracy for agiven power budget and number of frequency points requiredto resolve multi-pole impedance is discussed in detail. Finally,Section VII concludes the paper summarizing the merits of theproposed technique for Bio-impedance measurement.

II. MEASUREMENT SETUP

The measurement setup used in this study is shown in Fig. 1.In order to measure the Bio-impedance together with ECG

1932-4545 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

KUBENDRAN et al.: ERROR CORRECTION ALGORITHM FOR HIGH ACCURACY BIO-IMPEDANCE MEASUREMENT 197

Fig. 1. Bio-Impedance Measurement ASIC Block Diagram.

or EEG, there are 3 channels in the IC; 1 for ECG and 2 forimpedance (In- and Quadrature-phase channels). These signalsare separated after demodulation as shown in Fig. 1. Here, thedefault measurement frequency is 10 KHz [8].The square wave current is generated in the PCB using the

clock from the IC. The commercial level shifter is used forvarying the magnitude of the current. In total, four electrodes areconnected to the body – 2 for source and sink of the current to thebody, 2 for the signal measurement (both ECG and impedancetogether). One of them is used as a reference electrode whichis connected to a DC bias of around 600 mV for proper ampli-fication. The stimulation current is applied over the unknownimpedance to generate a voltage which passes through a HighPass Filter (see Fig. 1) and becomes the input to the ASIC. Thefront end stage of the ASIC comprises a Trans-Impedance am-plifier followed by a Tran-Conductance amplifier, which am-plifies the input voltage by a combined gain, . This signalis then demodulated using a Quadrature scheme to generatein-phase and quadrature-phase output voltages. The voltagesread from the ASIC are then amplified further by external gainstages of gain and respectively, as shown in Fig. 1. Themeasured voltages provide the inputs for the error correction al-gorithm proposed in this paper.

III. MATHEMATICAL ANALYSIS

The most common electrical tissue model is, where , are extra/intra-cellular

resistances and is membrane capacitance. However, inthis paper, an equivalent model as shown in Fig. 2 was usedwhich is more suitable for analysis [1]–[3]. This model hasa parallel combination of membrane resistance, andmembrane capacitance, in series with a base resistance,

. The purpose of this mathematical analysis is to calculateexact values for , and from output voltagesmeasured in the I-channel and Q-channel as shown in Fig. 1.

Fig. 2. Electrical model of tissue impedance.

Assuming a simple RC with a series model for thetissue, the tissue impedance to be measured is given by

(1)

Stimulation current is applied on the tissue and the complexvoltage developed across the tissue is measured using the frontend stage of the chip. The real and imaginary parts of the voltageacross the impedance are separately measured using quadraturedemodulation in two channels, In-phase (I) and Quadrature (Q)channels, respectively.

A. Sine Stimulation Current – Sine Demodulation Clock

In this analysis, fully differential bi-directional sinusoidalwave stimulation current and demodulation clock signals areconsidered, which can be expressed as follows:

(2)

where and are in-phase and quadrature-phase de-modulation signals with no physical units, i.e., they are dimen-sionless. and are the amplitudes of the stimulation currentand demodulation clock respectively.The input voltage signal to the ASIC, generated by the stim-

ulation current on the tissue impedance, is then amplified bythe front end stage by a gain , before demodulation. So theoutput of the in-phase demodulator can be derived as follows:

(3)

The demodulated output can be low pass filtered, using LPFwith cutoff frequency less than the modulation frequency, toobtain the dc level at the output, i.e., the cosine terms in (3) areremoved. Simplifying further, is given by

(4)


Note that, . So, theseries base resistance, can be estimated from the I-channelvoltage measured at the highest frequency in the given range.Similarly, the demodulated output of the Q-channel after

passing through the LPF is

(5)

The Error Correction Factor ( ), required by the algorithmdescribed in Section IV, is given by the ratio of demodulatedvoltages.

(6)

The impedance parameters can now be expressed in terms of, , , and , as described below

(7)

Thus the impedance parameters can be determined preciselyusing single tone sinusoidal waves with the least error.

B. Square Stimulation Current – Square Demodulation Clock

In this analysis, fully differential bi-directional square wavestimulation current and demodulation clock signals are consid-ered, which can be expanded using Fourier series and expressedas follows:

(8)

Similar to the previous case, the output of the in-phase de-modulator can be derived as follows:

(9)

The demodulated output can be low pass filtered to obtain thedc level at the output, i.e., the cosine terms in (9) are removed.Also, . Simplifying further

(10)

Note that, . So, theseries base resistance, can be estimated from the I-channelvoltage measured at the highest frequency in the given range.Now, the demodulated output of the Q-channel after passing

through the LPF is

(11)

In this case

(12)

Where is a scaling factor which is the ratio of the summa-tions in (10) and (11). This factor directly contributes to the errorin measuring and .Since the tissue model has a complex impedance of in

parallel with , the shape and phase of the voltage acrossthe tissue depends on the value of the RC time constant and thedominant component ( or ) at the measurementfrequency. For example, if is dominant, the voltage is asquare wave. And if is dominant, the voltage waveformis a triangular wave. The large error in measurement is due tothe higher order odd harmonic components which are translatedto dc level due to modulation and demodulation and hence dis-torting the measured voltages. This error can be reduced greatlyby using the parameter , defined in (12), since the inaccu-racy is highest around the pole frequencies ( )where the summations do not converge to a determined value.The measurement frequency can be swept to change the valueof with fixed and . The expressions forand are tabulated in the next section. The lower limit ofthe frequency range is determined by safety constraints and theupper limit is given by the amplifier bandwidth.

C. Sine Stimulation Current – Square Demodulation Clock

In this analysis, consider a fully differential bi-directionalsine wave stimulation current generated by a DAC of finite res-olution of bits and samples per cycle. Let be the time pe-riod of the sine wave and the sampling interval. The Fourier


series of reconstructed discrete sinusoidal waveform is derivedas follows:

So, ;Let be the peak-peak amplitude of the sine signal.

For a -bit resolution DAC, the LSB magnitude is givenby, . For a continuous sine wave signal,

, the discrete sampled signal is

(13)

where is the quantization noise component atthe sampling interval. Assuming a uniform distributionof the probability density function for the random process

, with zero mean, the upper bound of the magni-tude of the noise component is and the variance is .

The demodulation clock signals are now sampled squarewaves.

(14)

So, the input signal before amplification, , is given by

(15)

This signal is then amplified by a gain , before demodu-lation. So the input to the demodulator is

(16)

The output of the in-phase demodulator can be derived asfollows:

In order to compute , (13) was split into the signal anderror component. The signal component is then demodulated

similar to the sine-sine case. Before the error component is con-sidered, the minimum sampling rate for a given DAC resolutionneeds to be determined. The number of samples per cycle, hasto be large enough such that the error introduced due to sampletiming is less than the error introduced by finite DAC resolution.

(17)

Here, indicates the floor function to the nearest integer.The worst case was considered to determine the minimum sam-pling rate, ,

To get , substitute sampling instant time, where . Then take

limit , and approximate and

(18)The minimum sampling rate is given by

(19)

Given the sampling rate is higher than the limit determinedabove, the error component for a given sinusoidal stimulationcurrent and square demodulation clock can be determined pre-cisely. Fig. 3 shows graphically the product of and

which gives the demodulated error voltage waveform.It can be seen in Fig. 3 that at all sampling in-stants. The error maximum, i.e., occurs justbefore the sampling instants. The sign of the error component isbased on the fact that the reconstructed sine samples from DACwill always be less than the original sample since the DAC re-solves the value using the floor function.Now the demodulated voltage is obtained by simple multipli-

cation of the two waveforms as shown in Fig. 3. The resultantwaveform has a dc value since it is always positive. The demod-ulated output can be low pass filtered to obtain the dc level atthe output, i.e., the cosine terms are removed since they havehigher frequencies of multiples of . The dc error value ishalf of the peak amplitude as seen in Fig. 3.Now, demodulator output can be simplified as

(20)

Note that


Fig. 3. Voltage Error using Quantized sine wave for stimulation.

Similarly, the demodulated output of the Q-channel afterpassing through the LPF is

(21)

Here the parameter is given by,

We can see that the output voltage has signal and noise com-ponent. . In order to make the noise compo-nent negligible ( error), the magnitude of should beat least 100 times smaller than the signal to be measured, ,i.e., . This equation can be evaluated to find theminimum resolution of DAC.Neglecting the noise component

(22)

The impedance parameters can now be expressed in terms of, , , and , as described below

(23)

Minimum Resolution and Sampling Rate: In order to neglectthe noise components ( error), the minimum DAC resolu-tion is given by

TABLE IEVALUATION OF MODEL PARAMETERS FROM MEASURED VOLTAGES

Using (19), the minimum sampling rate is

IV. ALGORITHM FOR ERROR CORRECTION

A. Single/Multiple Frequency Measurement

The formulation in Table I based on the range of value of theparameter , gives the pseudo code for the algorithm that canbe implemented in Matlab and simulated for different values of

, , and measurement frequency. However, if onlyone measurement is done at a particular frequency, it is impos-sible to estimate all three tissue model parameters, i.e., ,

and . At best, the measurement will be accuratein the case when dominates when theimpedance reduces to a pure resistance, in series with

. In cases where and are comparable,the above algorithm can be directly extended for multiple fre-quency measurements. The infinite series sum in (10) and (11)are determined exactly in cases when or .In the third case, the sum is approximated to first 10 terms toget and . Initially, is taken as 1 since RC time constantis unknown. Once an estimate is obtained, the same Table I canbe reused with the value, in order to compute the exact valuesof and . is determined by applying mathemat-ical limits to the summations in (10) and (11). for

, for , else . For allcases, and .Though is theoretically evaluated at infinite frequency,

the expressions derived in Table I are valid even for finite rangeof measurement frequencies except that will be largerthan the actual resistance. In this case, the RC poles which areout of range (i.e., very small capacitance or resistance giving

) can be simplified as series resistance since thecapacitance can be taken as an open circuit. The series resis-tance is absorbed into making it larger. However, the RCpole frequencies within range remain the same and can be de-termined precisely using the algorithm.Assuming tissue impedance parameters do not change with

frequency, several measurements can be taken across decadesof frequencies, say 10 points per decade from 1 Hz to 1 MHz[11], it is possible to obtain a precise impedance variation plot


Fig. 4. Modified Electrical model of tissue impedance [13].

with frequency, i.e., , versus [12]. From this plot, thetissue model parameters can be accurately measured includingthe RC time constant.

B. Multiple Pole Multiple Frequency Measurement

Consider a more complex impedance model with a base re-sistance in series with number of parallel RC networks

formingmultiple poles (time constants) as shown in Fig. 4. This modelis derived in detail in [13] for deconvolution of electrochemicalimpedance. Though a different scenario, it is a generic modeland can be extended for tissue impedance too. The algorithmdescribed in Section IV.A with measurements at multiplefrequencies can be used for the generic impedance model. Theresulting plots of impedance variation with frequency can beanalyzed to evaluate the number of poles and their values.

V. RESULTS

A. Simulation Results

The algorithm described in Section IV was implemented inMatlab and several simulations were used to verify the accuracyof the algorithm in predicting the impedance model parameters.The values for the parameters were chosen for better illustrationof the algorithm and may not reflect actual tissue impedancevalue. The impedance graphs generated from a few examplesare given below,1) Case 1 – Single Pole: Consider a single pole impedance

with , and .Assume a stimulation current of amplitude 1 applied acrossthe tissue, 1 V amplitude modulation clock and an amplifier gainof 5000 before demodulation into I- and Q-channel voltages.The measurement frequency is swept from 1 Hz to 1 MHz, i.e.,6 decades with 10 points per decade [11], [14].At low frequencies when is dominant

( ), the effective impedance mea-sured will be purely resistive, in series with , witha value of . At higher frequencies when

is dominant ( ), again the impedanceis purely resistive and equal to . The polefrequency

(24)

2) Case 2 – Multiple Poles: Consider a two pole impedancewith , and ,

and . Assume the other con-ditions are the same as in Case 1.At low frequencies when is domi-

nant( ), the effective impedancemeasured will be purely resistive, in series with ,

with a value of . At very high fre-quencies when is dominant ( ), againthe impedance is purely resistive and equal to .The pole frequencies are

(25)

The algorithm implemented in Matlab takes these valuesas input parameters to evaluate ideal impedance and thenestimates the model parameters to obtain measured impedance.Using Table I, the values of and were computedat each frequency point. The poles were resolved by analyzingthe resistance and reactance curves. The pole resistance valuesare given by the difference in values of the flat regions inthe resistance curve. The pole capacitances are given by thelocal minima’s in the reactance curve, . Theestimated impedance magnitude and the corresponding errorplots for different demodulation methods are shown in Fig. 5.From the plots it can be clearly seen that the impedance valuescalculated at dc, high frequency, and the pole frequencies arematched accurately. The peak error reduces from 9.3% in theuncompensated case to 1.3% as seen in Fig. 6. Also impedancevariation with frequency after calibration is reproduced with anoverall average accuracy of for the single pole case,and for two poles.The simulation results for different cases are summarized in

Table II. From the Table II, the sine-sine case gives the bestresults whereas the square-square case without compensationgives the worst for both single pole and two pole RC networks.The sine-square case with gives better results than thesquare-square case but will end up consuming more power. Amore suitable alternative for low-power applications is using theerror correction algorithm for the square-square case which hasa slight decrease in accuracy compared to the sine-square case.The accuracy can be improved further by increasing the

frequency measurement points per decade, but will degradeas number of poles increases. Refer Section VI for furtherdiscussion. Hence the algorithm can be used for any genericimpedance model and not limited to Bio-impedance.

B. Measured Results

The algorithm described in Section IV was used for the mea-surement of impedance of saline solution with different concen-trations. The impedance across the AgCl gel electrodes, whichis primarily resistive, was measured using a standard ImpedanceAnalyzer and a PCBwith a test chip. The gain error in PCBmea-surements was initially calibrated using ideal components. Themeasurement clock frequency of the test chip was set to 10 KHz.The accuracy of in-vitro measurements using the calibrated testchip and post processed using the proposed algorithm, whencompared to that of the impedance analyzer, gives a worst caseerror of 0.72%. The results are tabulated in the Table III and theplot is presented in Fig. 9.The measured results for an impedance model as shown in

Fig. 2 using ideal components with values


Fig. 5. Impedance Magnitude Plot for a single pole tissue model.

Fig. 6. Impedance Error Plot for a single pole tissue model.

Fig. 7. Impedance Magnitude Plot for a two pole tissue model.

Fig. 8. Impedance Error Plot for a two pole tissue model.

and were used to evaluate the algorithm. Themeasurement frequency was swept from 10 KHz to 210 KHzand quadrature demodulated voltage readings, and wererecorded from a setup as discussed in Section II. The setup used

TABLE IISIMULATED RESULTS SUMMARY

TABLE IIIMEASUREMENT RESULTS FROM SALINE PHANTOM SOLUTION

Fig. 9. Measurement Accuracy with Distance between Electrodes.

a square wave stimulation current of amplitude, ,supply voltage , AFE gain setting of andexternal amplifier gain, and [8]. Usingthese values, and were first rescaled and calibrated.Then the parameter was computed at each measurementfrequency. From this curve, and using the proposed algorithm,extra-cellular resistance was calculated as(2.37% error). Also the intra-cellular resistance and mem-brane capacitance values were found to be( error) and ( error). Fig. 10shows the comparison between ideal and measured in-phaseand quadrature-phase voltages.Table IV shows a comparison of measurement error with

other works using different techniques for error correctionlike high resolution current meter [15], double pulse injection(voltage pulse followed by a current pulse) [16], multiplefrequency stimulation current [17] and Least Mean Squarealgorithm [18]. From Table IV, it can be seen that the error,after using the algorithm, is comparable in most cases, butsometimes larger. These works report error in measuringresistance and capacitance but they do not mention power


Fig. 10. Measured versus Ideal In-phase and Quadrature phase voltages.

TABLE IVCOMPARISON WITH OTHER WORKS

consumption. A more effective method should leverage onaccuracy, complexity of implementation and power consumed.So a Figure of Merit (FoM) is proposed in this paper whichindicates the best tradeoff (refer Section VI).

VI. DISCUSSION

Various tradeoffs are involved in choosing one of the aboveproposed methods for high accuracy Bio-impedance measure-ment in low power healthcare applications. The primary tradeoffis between power consumption and measurement accuracy asa function of the DAC resolution. The square-square stimula-tion-demodulation method is a special case where the DAC res-olution . Similarly, the pure sine-sine method is the casewhen the resolution is infinite, i.e., . In order to quantifythe effectiveness of the algorithm for a given resolution of theDAC, a Figure of Merit (FoM) can be proposed as follows:

(26)

Where is the supply current and LMSE is the Loga-rithmic Mean Square Error in measurement. Consider a currentsteering binary DAC architecture with -bit resolution.

(27)

Fig. 11. Figure of Merit versus DAC Resolution.

Where is the unit cell current and is the cur-rent consumed by other peripheral blocks in the DAC – bias,decoder, latches etc. Assuming constant , only the supplycurrent is a function of DAC resolution for sine-square case.Also the error is given by the ratio from (20) and(21) for and . Logarithmic error was used since the errordecreases several orders of magnitude as DAC resolution in-creases ( to ). The mean square of this logarithmicerror over frequency range of measurement was used to evaluatethe overall accuracy of the method. Fig. 11 shows plots of FoMas a function of the DAC resolution for both uncompensatedcase and using the proposed error correction algorithm.was taken as 10 and as 100 . FoM was scaled bya factor of 1000 for simplicity.From (26) and (27), as increases, FoM will decrease in-

versely. Similarly, as error increases on a logarithmic scale, FoMwill decrease inversely to the error squared. For lower DAC res-olution, is small but error is high. For higher resolution,is large but error is low. Since and ,FoM will be maximum at an optimal resolution for the DAC,

. Hence, the product helps us determine the optimal DACresoultion (which gives highest FoM) for tradeoff betweenand LMSE. As seen from Fig. 11, increasing DAC resolution be-yond theoptimal gives lesser benefit.Also, it canbenotedthat this value is less than computed in Section III, sincethe previous case had a constraint only onmeasurement error buthere power consumptionwasalso considered.For square-square case with/without compensation, and

LMSE are independent of DAC resolution as shown in Fig. 11.But FoM can be computed assuming a 1 bit DAC and errorcomputed from Table I. From Fig. 11 it is clear that, FoM forsquare-square without compensation is less than sine-square formost cases but with compensation, it is significantly higher thanall the cases. Sine-square can also be compensated to maximizeFoM but still it is lesser compared to square-square compen-sated case since is larger for the sine-square case thougherror is almost equal for optimal (see Fig. 6). Specifi-cally, FoM is improved for low resolution cases using the algo-rithm, which is highly advantageous for low power biomedicalapplications. This shows the error correction algorithm givesbest tradeoff with power consumption and measurement accu-racy for both sine-square and square-square cases.Figure of Merit of this work could not be compared with

others in Table IV, since power consumed by the systemwas not


provided. However, since the techniques for error correction inother works require a more complex implementation (high reso-lution current meter, double pulse injection, multiple frequencystimulation current, LMS algorithm etc), it would be safe to as-sume that the FoM will be better for the proposed algorithmsince the error numbers are comparable and our method willconsume less power by using only square waves.The other tradeoff is between number of frequency points

per decade and number of poles in the unknown impedance.Impedance measured is constant at dc, very high frequenciesand frequencies where there are no poles nearby. However,around poles the number of frequency points need to be higherto measure impedance variation accurately. In addition, thenumber of poles in the unknown impedance determines the totalnumber of frequency points needed in the given range of mea-surement. Finally, it can be noted that though Bio-impedancecan be best modeled as a distributed RC network, a lumpedapproximation might be sufficient to derive the informationneeded. As a rule of thumb, 10 log-spaced frequency pointsper decade are required to resolve the pole effectively withinone-tenths of a decade.It is important to emphasize the need for accurate post calibra-

tion in order to remove any offset errors, gain errors and phasemismatch errors contributed from themeasurement bench setup.

VII. CONCLUSION

Bio-impedance measurement using different stimulation-de-modulation techniques have been analyzed in detail in thispaper. While using single tone sinusoidal waves providethe best accuracy, they are difficult to generate on-chip andconsume more power. In contrast, square waves are easy togenerate but result in high inaccuracy due to the higher har-monic terms demodulated to dc.An error correction algorithm is proposed for the case using

square waves to improve themeasurement accuracy. To be com-prehensive, a generic case has been considered with sine wavegenerated using a finite resolution DAC and measurement accu-racy derived. The sine-sine and square-square cases are shownto be a subset where the DAC resolution is 1 for square wavesand for sine waves. A figure of merit has been proposed toclearly show the tradeoff between power consumption and mea-surement accuracy. Finally the error correction algorithm pro-posed for the case using square waves is shown to improve theFoM without compromising on power consumed.Thus, Bio-impedance can be measured using square wave

stimulation current along with the proposed algorithm for errorcorrection. This method has the advantage of lower powerconsumption compared to sine wave demodulation with onlya slight decrease in accuracy. The error in measurement isshown to be compensated to a great extent with the proposedalgorithm. Hence this method can be used in physiologicalsignal monitoring applications using quadrature modulation forsignal acquisition.

REFERENCES

[1] S. Grimnes and Ø. G. Martinsen, Bioimpedance & Bioelectricity Ba-sics, 2nd ed. London, U.K.: Academic, 2000.

[2] E. T. McAdams and J. Jossinet, “Tissue impedance: A historicaloverview,” Phys. Meas., vol. 16, pp. A1–A13, 1995.

[3] J. J. Ackmann, “Complex bioelectric impedance measurement systemfor the frequency range from 5 Hz to 1-MHz,” Ann. Biomed. Eng., vol.21, pp. 135–146, 1993.

[4] R. F. Yazicioglu, S. Kim, T. Torfs, H. Kim, and C. VanHoof, “A 30analog signal processor ASIC for portable biopotential signal moni-toring,” IEEE J. Solid-State Circuits, vol. 46, no. 1, pp. 209–223, Jan.2011.

[5] R. C. Kubendran, S. Kim, and R. F. Yazicioglu, “Error correction al-gorithm for high accuracy bio-impedance measurement in wearablehealthcare applications,” in Proc. IEEE Int. Symp. Circuits and Syst.,May 2013, pp. 1292–1295.

[6] H. Kim et al., “A configurable and low-power mixed signal SoC forportable ECG monitoring applications,” IEEE Trans. Biomed. CircuitsSyst., 2013, Early Access, IEEE Xplore.

[7] N. Van Helleputte, S. Kim, H. Kim, J. P. Kim, C. Van Hoof, and R.F. Yazicioglu, “A 160 biopotential acquisition IC with fully inte-grated IA and motion artifact suppression,” IEEE Trans. Biomed. Cir-cuits Syst., vol. 6, no. 6, pp. 552–561, Dec. 2012.

[8] S. Lee et al., “A low-power and compact-sized wearable bio-impedance monitor with wireless connectivity,” in Proc. J. Phys,Conf. Ser., 434 012013.

[9] M.Min, T. Parve, V. Kukk, and A. Kuhlberg, “An implantable analyzerof bio-impedance dynamics: Mixed signal approach,” IEEE Trans. In-strum. Meas,, vol. 51, no. 4, pp. 674–678, Aug. 2002.

[10] A. Yúfera et al., “A tissue impedance measurement chip for myocardialischemia detection,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 52,no. 12, pp. 2620–2628, Dec. 2005.

[11] B. Rigaud, L. Hamzaoui, M. R. Frikha, N. Chauveau, and J. P.Morucci, “In vitro tissue characterization and modelling using elec-trical impedance measurements in the 100 Hz – 10 MHz frequencyrange,” Phys. Meas., vol. 16, pp. 15–28.

[12] A. Ronk, M. Min, and T. Parve, “Simultaneous multi-frequency bio-impedance measurement applying synchronised uniform or non-uni-form sampling,” in Proc. IFMBE 17, 2007, ICEBI, SpringerLink, pp.185–189.

[13] H. Schichlein, A. C. Muller, M. Voigts, A. Krugel, and E. Ivers-Tiffee,“Deconvolution of electrochemical impedance spectra for the identifi-cation of electrode reaction mechanisms in solid oxide fuel cells,” J.Appl. Electrochem., vol. 32, pp. 875–882, 2002.

[14] B. Sanchez and R. Bragos, “Fast electrical impedance spectroscopy formoving tissue characterization using bilateral quasilogarithmic multi-sine bursts signals,” in Proc. IFMBE 22, 2008, ECIFMBE, Springer-Link, pp. 1084–1087.

[15] J. Lario-Garcýa and R. Pallàs-Areny, “Measurement of three indepen-dent components in impedance sensors using a single square wave,”Sens. Actuators A, pp. 164–170, Aug. 2003.

[16] A.Mejía-Aguilar and R. Pallàs-Areny, “Electrical impedancemeasure-ment using voltage/current pulse excitation,” in Proc. 19th IMEKOWorld Congr., Fundamental and Appl. Metrology, Sep. 2009.

[17] T. Fukumoto et al., “Temporal resolution of the skin impedance mea-surementin frequency-domain method,” IEEE Trans. Biomed. Eng.,vol. 54, no. 1, pp. 170–173, Jan. 2007.

[18] J. Diao et al., “High accuracy biological impedance measurementsystem design and calibration,” in Proc. 3rd Int. Conf. Digital Manu-facturing Automation, Aug. 2012, pp. 466–470.

Rajkumar Kubendran (S’10–M’12) was born inMadurai, India, in 1987. He received the B.Tech.degree from the National Institute of Technology,Tiruchirappalli, Tamil Nadu, India, and the M.S.degree in electrical and computer engineering fromPurdue University, West Lafayette, IN, USA, in2008 and 2012, respectively.From 2010 to 2011, he was a Research Assistant

working on wireless RF powering circuits forbiomedical implants. During January-May 2012,he was an International Visiting Scholar at imec,

Leuven, Belgium, where he did an internship on ECG readout circuit designand application. Since August 2012, he has been a RFIC Design Engineerat Qualcomm Inc., Boxborough, MA, USA. His academic interests includelow power analog and RF design, and high accuracy measurement techniquesfor biomedical applications. He received the Best Student Paper Award atISCAS 2013 in Beijing, China, for his work on error correction algorithm forbio-impedance measurement.


Seulki Lee (S’07) received the B.S., M.S., and Ph.D.degrees fromKAIST, Daejeon, South Korea, in 2007,2009, and 2012, respectively, all in electrical engi-neering.Her doctoral research focused on low power IC de-

sign and wearable system development for wearablebody area network (WBAN), such as a sleep moni-toring system. In March 2012, she joined the HolstCentre imec, Eindhoven, The Netherlands, as a Re-searcher on the body area network (BAN) team. Hercurrent research interest mainly focuses on the low

power bio-impedance system and application.

Srinjoy Mitra (M’09) received the B.S. degreefrom the University of Calcutta, Kolkata, India,the M.Tech. degree from the Indian Institute ofTechnology, Bombay, India, and the Ph.D. degreefrom the Institute of Neuroinformatics, Universityof Zurich, ETH Zurich, Zurich, Switzerland.He briefly worked in the microelectronic industry.

He spent two years as a Postdoctoral Researcherat The Johns Hopkins University, Baltimore, MD,USA, and joined the medical electronics team atimec, Heverlee, Belgium, in 2010. His interests are

in low-power analog circuits for biomedical interfaces.

Refet Firat Yazicioglu (M’06) received the B.S. andM.S. degrees in electrical and electronics engineeringfrom Middle East Technical University (METU),Ankara, Turkey, in 2001 and 2003, respectively,and the Ph.D. degree in electronics engineering,with an emphasis on the design of low-power andlow-noise readout circuits for portable biopotentialacquisition systems, from Katholieke UniversiteitLeuven, Leuven, Belgium, in 2008.Currently, he is the R&D Team Leader at imec,

Leuven, Belgium, where he is leading the activitieson analog integrated circuit design for portable and implantable biomedical ap-plications. During his research, he has coauthored more than 70 publications,three book chapters, and a book on ultra-low-power circuit and system designfor biomedical applications. He has developed several generations of integratedcircuits for wearable and implantable healthcare applications.Dr. Yazicioglu was the co-recepient of the Best Student Paper Award at the

IEEE International Symposium on Circuits and Systems (ISCAS) 2013, IEEEBiomedical Circuits and System Conference (BioCAS) 2011, Smart SystemsIntegration (SSI) Conference, 2008, and Sensors and Transducers Journal 2008.He serves on the technical program committees of European Solid State Circuits(ESSCIRC) and International Solid State Circuit Conference (ISSCC). He wascochair of the Biomedical Circuits and Systems Conference (BioCAS) 2013 inRotterdam, The Netherlands.

Error Correction Algorithm for High Accuracy Bio-Impedance Measurement in Wearable Healthcare...

Documents

Transcript of Error Correction Algorithm for High Accuracy Bio-Impedance Measurement in Wearable Healthcare...