{"title": "A Micropower CMOS Adaptive Amplitude and Shift Invariant Vector Quantiser", "book": "Advances in Neural Information Processing Systems", "page_first": 671, "page_last": 677, "abstract": null, "full_text": "A Micropower CMOS Adaptive Amplitude and \n\nShift Invariant Vector Quantiser \n\nRichard J. Coggins, Raymond J.W. Wang and Marwan A. Jabri \n\nComputer Engineering Laboratory \n\nSchool of Electrical and Infonnation Engineering, J03 \n\nUniversity of Sydney, 2006, Australia. \n\n{richardc, jwwang, marwan} @seda1.usyd.edu.au \n\nAbstract \n\nIn this paper we describe the architecture, implementation and experi(cid:173)\nmental results for an Intracardiac Electrogram (ICEG) classification and \ncompression chip. The chip processes and vector-quantises 30 dimen(cid:173)\nsional analogue vectors while consuming a maximum of 2.5 J-tW power \nfor a heart rate of 60 beats per minute (1 vector per second) from a 3.3 V \nsupply. This represents a significant advance on previous work which \nachieved ultra low power supervised morphology classification since the \ntemplate matching scheme used in this chip enables unsupervised blind \nclassification of abnonnal rhythms and the computational support for low \nbit rate data compression. The adaptive template matching scheme used \nis tolerant to amplitude variations, and inter- and intra-sample time shifts. \n\n1 \n\nINTRODUCTION \n\nImplantable cardioverter defibrillators (ICDs) are devices used to monitor the electrical \nactivity of the heart muscle and to apply appropriate levels of electrical stimulation if ab(cid:173)\nnonnal conditions are detected. Despite the considerable success of ICDs they suffer from \na number of limitations including an inability to detect and treat some abnonnal heart \nrhythms and limited data recording capabilities. \n\nWe have previously shown that micropower analogue Multi-Layer Perceptron (MLP) neu(cid:173)\nral networks can be trained to separate such arrhythmia [4]. However, MLPs are best suited \nto learning the boundary between classes whereas a vector quantization scheme allows a \nmeasure of the probability density of the morphological types to be estimated. \n\nMany analogue vector quantiser (VQ) chips have been reported in the literature. For ex(cid:173)\nample, a 16x256 500 kHz 50 mW 2 J-tm CMOS vector AID converter [10] and a 16 x 16 \n300 kHz 0.7 mW 2 J-tm CMOS analogue VQ [1]. These correspond to an energy per match \n\n\f672 \n\nR. J. Coggins, R. J. W. Wang and M. A. Jabri \n\nper dimension of 24 pI and 9 pI respectively. The integrated circuit (lC) described in \nthis paper is distinguished from these approaches in that it is specifically targeted for the \nlow power, low bandwidth application of ICEG classification and compression. Our chip \nachieves vector matching (without the winner take all function) to 7 bit 30 dimensional \nvectors with three coefficient linear prediction, at an energy consumption of 15 pI per tem(cid:173)\nplate per dimension using a 1.2 pm CMOS process. Although this figure is greater than \nthat for [1] it should be noted that in [1] the mean absolute error metric is used rather than \nthe squared Euclidean distance and no provision is provided for linear transformation of \nthe incoming analogue vector. \n\n2 ADAPTIVE DATA COMPRESSION \n\nRecording of ICEGs in ICDs is currently very limited due to the amount of memory avail(cid:173)\nable and the power/area cost of implementing all but the simplest compression techniques. \nMicropower template matching however, enables large amounts of the signal to be encoded \nas template indices plus amplitude parameters. Effective compression ofthe ICEG requires \nadaptation to the short term non-stationary behaviour of the ICEG [2] . In particular, short \nterm amplitude variations, lag variation, phase variation and ectopic beats (which origi(cid:173)\nnate from the ventricles of the heart and have differing morphology) reduce the achievable \ncompression. The impact of ectopic beats can be reduced by increasing the number of \ntemplates. This can often be achieved without increasing the code book search complexity \nby using associated timing features. The amplitude and shift variations require short term \nadaptation of the template matching in order to minimise the residual error and hence raise \nthe compression ratio at fixed distortion. \n\n2.1 Amplitude and Shift Invariant Matching \n\nIn order to facilitate analogue implementation, a backward prediction procedure is used \nrather than the usual forward prediction [8]. This approach allows the incoming analogue \ntemplate to be manipulated in the analogue domain for amplitude and shift invariance pur(cid:173)\nposes. Consider the long term backward prediction problem described by, \n\n) b {x(n + a + 1) - x(n + a - I)} \n\n2 \n\n() -() b ( \n\nrb n = x n - OX n + a - I \n\n(1) \nwhere rb (n) denotes the backward residuals, x is a template which is a function of previous \nbeats, x( a) is the sampled ICEG signal, a the time index, n is the template index and bo and \nbl are the amplitude and phase coefficients respectively. bo scales the current beat to match \nthe template and hence is an amplitude term. b1 scales the central difference of the current \nbeat and is a function of the amplitude and phase corrections required to minimise the \nresiduals. To see why this is a phase term consider the Taylor expansion of Ax(t + \u00a2) to \nthe first derivative term around t, \n\nAx(t + \u00a2) = Ax(t) + A\u00a2x' (t) \n\n(2) \nwhere \u00a2 is a small phase shift of x(t) and A is the amplitude factor. When \u00a2 is due to \nsampling jitter then, -t :::; \u00a2 :::; t, where T is the sampling period. Provided that x(t) is \nsampled according to the Nyquist criterion, \u00a2 is sufficiently small for the first derivative \nterm to adequately account for the sampling jitter. Hence, bi accounts for the residual error \nremaining after optimisation of the integer a. a is approximately determined by the beat \ndetector of the ICD which attempts to detect the fiducial point of heart beats using filters \nand comparators. bo and b1 can be determined by minimising the squared error between \nthe current signal window and the previously recorded template which in this case has a \nclosed form solution in terms of correlation coefficients. However, in Section 3 we present \nan alternative iterative procedure suited to low-power analogue implementations. \n\n\fA Micropower CMOS Adaptive Amplitude and Shift Invariant Vector Quantiser \n\n673 \n\n3 SYSTEM ARCHITECTURE & IMPLEMENTATION \n\n'-\n\n\"\"~fC\"'-1 \u2022\u2022 \n\n... ~) \u2022\u2022\u2022 \"'J \n\nn. \nLL \n\nr , .. ::;-\n\n'.I~ \n\n1'11':::--\nI'U =-\n\nFigure I: Left: Block diagram of the adaptive linear transfonn VQ chip. Middle: Floorplan \nof the chip. Right: Photomicrograph of the chip. \n\nThe ICEG is first high pass filtered to remove the DC and then is bandpass filtered to \nprevent aliasing and enhance the high frequency component for beat detection. (This is \nthe filtering approach already existing in an ICD and therefore not implemented by us). \nThis then feeds the discrete time analogue delay line, which is continuously sampling the \nsignal at 250 Hz. The analogue samples are then transfonned by a two layer network. The \nfirst layer implements the linear prediction by adjusting the amplitude bo and the phase \nof the analogue vector. Note that the phase consists of two components, the coarse part \na corresponding to sample lags and the fine part b1 corresponding to intra-sample lags. The \nsecond layer calculates the distance between the linearly predicted vector and the template \nwen) to be matched. A comparator is provided so that a match to within a given threshold \nmay be detected. \n\n3.1 Chip Architecture \n\nInput to the IC is via a single analogue channel which is sampled by a bucket brigade \ndevice of length 30. The resultant 30 dimensional analogue vector is adaptively linear \ntransfonned to facilitate a shift and scale invariant match to a digital (7 bit per dimension) \ntemplate. The IC generates digital representations of the square of the Euclidean distance \nbetween the transfonned analogue vector and the digital template. A block diagram of the \nIC appears in Figure I. The IC has been fabricated. Perfonnance figures in this paper are \nbased on measurements of the chip fabricated in a 1.2/-Lm CMOS MOSIS process. \n\nThe block diagram shows the input signal being sampled by the bucket brigade device \n(BBD)[4]. The signal is sampled at a rate of 250 Hz. Existing circuitry in the defibrillator \ndetects the peak of the heart beat and hence indicates a coarse alignment (due to detection \njitter) to the template stored in the template DACs (TDACs). The BBD continues to sample \nuntil the coarse alignment is attained at which point the IC is biased up. The BBD now \ncontains a segment of the ICEG corresponding to one heart beat. The digital error output is \nthen monitored with the linear transfonn blocks configured to I: I mappings until an error \nminimum is detected indicating optimal sampling alignment. The three linear transfonn \ncoefficient DACs (CDACs) which are common to the 30 linear transfoqn blocks may then \nbe adapted to further reduce the matching error. The transfonnation can be represented by \nyen) = aox(n - 1) + alx(n) + a2x(n + 1) where ao corresponds to CDACO etc. This \nconstitutes a general linear long tenn prediction [8]. Constraining CDACO and CDAC2 to \nbe equal magnitudes and opposite signs results in a minimisation of errors due to phase \nand amplitude variation and a simpler adaptation procedure. The matching error is com(cid:173)\nputed via the squarer blocks and the summing node. The matching error consists of both a \nmagnitude and exponent thereby increasing the dynamic range of the error representation. \n\n\f674 \n\nR. J. Coggins, R. J. W Wang and M. A . Jabri \n\nThe magnitude is the output of the squarer block. The exponent is determined by control \nof a current reference in the squaring circuit. A reference DAC and precision current com(cid:173)\nparator provide the means of successive approximation AID conversion of the matching \nerror current [ERR. Using this scheme heart beat morphology can be classified by loading \ndifferent templates (TDAC values). A stream of beats may be compressed by identify(cid:173)\ning matches with continuously updated representations of previous beats. Close matches \nare encoded by an index and an amplitude coefficent while poor matches are encoded by \nquanti sed residuals which have been minimised by the linear prediction. \n\n3.2 Adaptation and Learning \n\nThe first step in the learning process is to determine a, the coarse phase lag. This can be \nachieved by shifting the delay line and evaluating the error until a minimum is reached. \nOnce the coarse phase lag a has been determined the error function to be minimised to \ncompensate for amplitude and phase variations is given by E = E~I (bOXi+bI~Xi-Wi)2, \nwhere the subscript i implicitly incorporates the coarse phase a. This is a quadratic in \nbo and bl . bo and bi can be optimised separately provided cross terms in E are negligible. \nHere the cross terms are given by E~I 2bobIXi~Xi = bobI(XN+IXN - XIXO). Thus, if \nthe end points of the N point window have approximately the same value (as is usually \nthe case for ICEG beats) then the cross terms in E are negligible and bo and bi can be \noptimised separately. \n\nSo the only remaining issue is how to optimise a single parameter. A simple linear search \ntakes at most 2b evaluations of E where b is the number of bits. A search based on bisection \ntakes b + 2 evaluations. Techniques involving gradient descent and conjugate gradient \nlead to more complex learning logic with minor reductions in the number of evaluations. \nTherefore, bisection is the best compromise between the number of evaluations and the \ncomplexity of the learning state machine. \n\nOnce the best template match has been achieved, learning may also then be applied to \nthe template itself depending on the application and context. For example, in the case \nof adaptive classification a weight perturbation algorithm [6] could be used to adapt the \ntemplate for morphological drift based on heart rate information. Similarly, for a data \ncompression application, if the template match exceeds a fidelity criterion the template \nmay be adapted and the template changes logged in the compression record. \n\n3.3 Building Blocks \n\nIn order to implement the template matcher, sub-threshold analogue VLSI building blocks \nwere designed. All transistors in the building blocks operate in weak inversion exclusively. \nWe do not have the space to describe all of the building blocks, so we will focus here on \nthe linear transform and squarer cells. \n\n3.3.1 Linear Transform Cell \n\nThe linear transform (LT) cell consists of three linearised differential pairs [7] with their \nbiases controlled by the coefficient DACs (CDACs) (see Figure 2(a\u00bb. The nature ofthe lin(cid:173)\nearisation is controlled by the ratio of the aspect ratios.ofM3 to M5 and M4 to M6. Methods \nfor choosing this ratio are discussed in [5]. Denoting the aspect ratio of a transistor by S we \nchose S3/ S5 = S4/ S6 = 4. This introduces some ripple in the transconductance while \nincreasing the asymptotic saturation voltage to 4nUT compared to nUT for the ordinary \ndifferential pair. Signed coefficients are achieved by switches at the outputs of the differ(cid:173)\nential pairs. The template DACs (TDACs) have differential outputs to form the difference \ny(n) - w(n) where w(n) is the nth template value. \n\n\fA Micropower CMOS Adaptive Amplitude and Shift Invariant Vector Quantiser \n\n675 \n\n3.3.2 Squaring Cell \n\nThe squaring function must meet the following design constraints. It should have current \ninputs and outputs in order to avoid linear current to voltage conversion at low currents. The \nsquared current must be normalised to the original linear range to avoid excessive power \nconsumption. The squaring function should avoid the MOS square law approach in order \nto conserve space and power, and the the available voltage range should be 3.3 V rail to \nrail. \n\nRCLK1 D--~--*\"-