Development of an Efficient Algorithm for Fetal Heart Rate Detection: A Hardware Approach

Size: px

Start display at page:

Download "Development of an Efficient Algorithm for Fetal Heart Rate Detection: A Hardware Approach"

Roderick Clarke
5 years ago
Views:

1 Developmen of an Efficien Algorihm for Hear Rae Deecion: A Hardware Approach M. I. IBRAHIMY, M. B. I. REAZ 2, M. A. MOHD ALI 3, T. H. KHOON 2, A. F. ISMAIL Kulliyyah of Engineering, Inernaional Islamic Universiy, 5300 Selangor, Malaysia 2 Faculy of Engineering, Mulimedia Universiy, 6300 Cyberjaya, Malaysia 3 Faculy of Engineering, Universii Kebangsaan Malaysia, Bangi, Malaysia Absrac: - An algorihm has been developed for he simulaneous measuremen of he feal maernal hear raes from he maernal abdominal elecrocardiogram during pregnancy labor for feal monioring. The algorihm is based on cross-correlaion, adapive hresholding saisical properies in he ime domain. The algorihm was iniially developed simulaed in Visual C++. Once he funcionaliy is verified, i is hen convered in VHDL - hardware descripion language for FPGA implemenaion. The design is synhesized fied ino Alera s Sraix EPS0 using he Quarus II plaform because of is enhanced DSP capabiliy. Tes case resuls showed an error percenage of around ±0.3% ±0.5% for he deecion of maernal feal hear rae respecively. Key-Words: - Hear Rae, Elecrocardiogram, VHDL, FPGA Inroducion The elecrocardiogram (ECG) is he elecrical signal produced by he hear conains he disincive shape known as he QRS complex. The ime beween wo successive R peaks of he QRS complex is known as he RR inerval he hear rae (HR) is he reciprocal of he RR inerval expressed in Bea Per Minue (BPM). Elecronic feal hear rae (FHR) monioring is used o deermine if he feus is free from any complicaions such as anenaal ueroplacenal insufficiency feal hypoxia, o deermine he feal healh []. Coninuous effors are being made o produce more efficien accurae mehods o exrac he feal RR inerval, gain a beer inerpreaion of he FHR paerns by researchers. A presen, Doppler ulrasound has become a popular echnique of monioring he FHR abdominally bu aemps o produce a porable sysem have no been successful because of is sensiiviy o movemens [2]. Mehod uilizing he abdominal elecrocardiogram (AECG) has a beer prospec for long-erm monioring bu requires much signal processing o be done [2][3]. This mehod is non-invasive has poenial o convey he elecrophysiological informaion, which helps o deermine he condiions of he feus such as sress acidosis, uerine aciviy [3]. A beer single-lead mehod [4] has been adoped improved o exrac he maernal feal QRS complexes from he AECG. 2 Mehods 2. QRS Deecion The deecion of maernal QRS complexes is begun wih cross-correlaing he signal wih an average maernal QRS emplae. The cross-correlaion oupu of he signal x a each insan n wih he emplae s(k) is given by where M y( n) = h( k) x( n k). () k = 0 s( M k), h( k) = { 0, 0 k M elsewhere The emplae s(k) wih (M + ) equally spaced poins over 80 ms has been empirically found o be opimized for he deecion of maernal QRS complexes when M = 8. The widh of he emplae is based on he normal widh of he maernal QRS complex [5]. The emplae is coninuously updaed wih he deecion of R peaks o ake ino consideraion he variaion of shape of he maernal QRS complexes in AECG. The local maxima search rouine measures he slope of he cross-correlaed oupu by ' y ( n) = y( n) y( n ). (2)

2 assumes a maximum a sample (n - ) when he slope changes from y (n-) 0 o y (n) < 0. If no maximum is found in he subsequen 20 ms (assumed o be he minimum feal QRS duraion [6]), he sample value y(n-) corresponding insan are saved as he local maximum. This 20 ms search inerval is necessary o avoid aking small spikes on he slopes of he QRS complexes as maxima. Three values, V M > V M2 > V M3, heir ime insans corresponding o he larges hree local maxima are sored wihin an R wave search inerval. The lengh of he search inerval is iniially one second (in fac 024 ms for compuaional simpliciy) i is hen coninuously updaed afer he firs RR inerval measuremen. The one-second search inerval he saving of 3 local maxima assume ha he maernal hear rae (MHR) does no exceed 20 BPM which means a mos 2 maernal R peaks can be found in he iniial search inerval. If V M is validaed as he R peak hen he value V M2 is aken as he noise. V M3 is kep for cases when V M2 is validaed as he R peak. The hreshold used in he deecion is se iniially by assuming a minimum maernal R peak of 0 µv [6] i is coninuously updaed based on he levels of boh R peak noise. A possible maernal R peak is assumed o be found when he value V M exceeds his hreshold. V M2 is also considered as an R peak if he value is comparable o ha of V M he resuling hear rae is below 20 BPM, as earlier assumed. Hence he crieria: 2V M2 > V M. (3a) 52 ms. (3b) M2 M > If V M2 also exceeds he hreshold, he QRS emplae is compared wih he complexes associaed wih boh V M V M2. The one wih he leas mean square error is aken o be he R peak. The oher peak is assumed o be a large spike in he signal is posiion is saved for use in he feal R peak validaion rouine. If V M2 has he larger error, is posiion is saved only if inequaliy in Equaion (3a) applies, because smaller V M2 may be associaed wih an acual feal R peak. The running average used in his algorihm is performed o average he QRS emplaes, RR inervals, levels of R peak noise. The b-h value of he running average A(b) is given by a weighing of he previous average A(b-) plus ha of he new value C(b) as shown in he following equaion: where { k( b) } A( b ) k( b) C( ) A ( b) = + b. (4) k ( b ) = { b B,, b B b > B The running averages of noise R peaks (A N A R ) are esimaed over B recen values where, B = 8 in Equaion (4) has been empirically found o be effecive. Based on hese averages, wo hresholds, TM TM 2 are used in he R wave search. They are given by AR AN TM = AN + 4. (5a) TM TM 2 = 2. (5b) The adapaion of he hreshold o varying R peak noise levels, he R wave search inerval are based on he mehod proposed in [7]. If he maximum search limi is reached while he local maximum V M has a value less han TM, hen V M is aken as a possible R peak if i exceeds he second hreshold, TM 2. If no such V M is found, a signal loss is assumed. The local maxima values are hen se o zero for he subsequen R wave search. Four laes maernal RR inervals are mainained in record for he purpose of checking coincidences of he maernal wih he feal R waves. 2.2 QRS Deecion The maernal elecrocardiogram (MECG) complex is hen subraced upon deecion of a maernal QRS o remove he maernal conribuion from he abdominal signal. This complex is of fixed duraion, 60 ms before 320 ms afer he maernal R peak insan. This duraion assumes ha he average MHR is less han 25 BPM i should normally include he P T waves, if any. The MECG emplae is mached wih acual MECG in he abdominal signal by scaling i wih he facor, Value K = Value2 (6) where, Value < Value2. These values are obained from he cross-correlaion of abdominal signal wih maernal emplae auo correlaion of he maernal emplae. If he cross correlaion value is greaer han he auo correlaion value, hen he abdominal signal is muliplied by he facor K MECG emplae is subraced, if no, MECG emplae is muliplied by facor K subraced from he abdominal ECG signal.

3 The deecion of he feal QRS complex is begun wih differencing of local maxima minima on he oupu of he subraced signal when he ime marker coun, which was iniiaed a he second acceped maernal R peak, has reached 2048 ms. This duraion ensures ha he 2 second delayed samples are already wihin he MECG subraced region of he signal. Observing he waveforms, i is possible o differeniae beween feal evens noise even if he ampliudes are similar. This is parly because of he rapid large deflecions beween a local maximum he following local minimum when a feal bea has occurred. From Equaion (2), a minimum is assumed a sample (n - ) when he slope changes from y (n-) < 0 o y (n) 0. The absolue value of he difference beween successive peak valley is compued for each max-o-min inerval. The local maxima search rouine is performed on he oupu of he differencing of local maxima minima rouine, hree larges maxima, V F > V F2 > V F3 are kep as before. The iniial search inerval is 640 ms so ha a mos wo feal R peaks can be found by assuming he FHR does no exceed 87 BPM during he iniial search inerval. The firs search is repeaed for anoher subsequen 640 ms if he larges local maximum, V F is concurren wih a maernal QRS complex V F2 is smaller han a hreshold or is also concurren. The hreshold used in he FHR deecion is se iniially by assuming a minimum feal R peak of 5 µv [6] i is coninuously updaed. The rouine is similar o ha for he maernal case bu uses he following crieria o accep V F2 as a possible feal R peak:.5v F2 > V F. (7a) 320 ms. (7b) F2 F > The second search is repeaed if he acceped firs feal R peak is found o be concurren wih a maernal QRS complex or if 2 V F3 > V F. (8) i.e. he signal is noisy wih all is hree local maxima having comparable values. The feal maernal QRS complexes are concurren if 64 ms. (9) F M < where F M are he feal maernal R peak insans, respecively. The range in Equaion (9) accouns for possible overlap of he wo complexes, which are assumed o have widhs of ms respecively. The overlap is checked by relaing he feal R peak insan o he four laes maernal RR inervals. The subsequen feal R wave deecion procedure is he same as ha for he maernal R wave using wo hresholds, TF TF 2 which are se as in Equaion (5), according o he running average of he R peaks noise wih B = 8 in Equaion (4). The deerminaion of he feal R wave search inerval is also based on he mehod proposed in [7]. The second hreshold, TF 2 is used when he maximum search limi is reached. A signal loss is assumed when no maximum exceeding he hreshold is found. When he second hreshold is used o idenify a feal R peak, he peaks are averaged wih B = 4 so ha he firs hreshold will quickly adap o he smaller signal. Afer a possible feal R wave is found, a coninuaion of he search for up o 220 ms is carried ou unless he maximum search limi is reached. This forward searching reduces he possibiliy of false R wave deecion wih he assumpion ha he hear rae does no exceed 270 BPM. Then he program branches o he validae updae rouines. The validae rouine firs checks if V F3 > TF. (0a).5V F3 > V F. (0b) where TF is he hreshold used o deec V F. These condiions mean ha he feal R peak was obained in a very noisy signal. Oherwise, similar checks are made wih V F2, where V F2 > TF. (a).5v F2 > V F. (b) also imply a noisy signal. If V F is he only maximum above he hreshold hen i is aken as a feal R wave. If V F2 also exceeds he hreshold, hen V F is checked for coincidence wih possible spikes by relaing is insan o he four maernal values which are kep in he record. The spike posiion, S he posiion, F in he signal associaed wih he local maximum, are compared for 40 ms. (2) S F < which allows for he difference in correlaion delay when obaining S F respecively. If V F is idenified as a large spike in he signal, hen V F2

4 V F3 are assumed o be he feal R peak he noise, respecively. Thresholds search inerval limis are updaed according o he procedure described earlier he local maxima values are hen se o zero for he subsequen R wave search. 2.3 Hardware Implemenaion The QRS deecion algorihm was iniially implemened in Visual C++ because i is simpler faser o verify he funcionaliy reliabiliy. Then, he algorihm was implemened in VHDL; where Alera s Quarus II version 4.0 is used as he plaform. As a resul, for VHDL implemenaion he algorihm has o be hough of as a srucural, behavioral physical version of he algorihm. The advanage of using Quarus II is ha he sysem could be synhesized ino a physically available FPGA, or he buil-in simulaion device models. The buil-in simulaion device models emulae he real device wih he acual iming power values. Thus, is performance in erms of iming, speed, power consumpion, oal logic elemen couns funcionaliy could be ascerained. Modificaions can be easily performed is impac o he physical implemenaion, especially in erms of iming funcionaliy can be immediaely known. Fig. shows a simplified block diagram of he implemenaion of he sysem. Basically, he sysem is caegorized ino hree main blocks, he common, maernal feal blocks. The common blocks, shared by boh maernal feal blocks are he Daa Inpu, Memory Regiser, Memory Iniializaion Main Conrol. blocks consis of he Iniializaion s (XCM, Inipar, Inipar2 IniSub ), RR Inerval, Validaion, Subracion Correlaion & Local Maxima. blocks consis of Iniializaion, Exracion, Correlaion, Local Maxima Search, Templae Updae Coincidence. The pins for he sysem PIN_ NEWDATA, PIN_DATA PIN_DATAREQ are used o inerface wih an exernal module o rerieve new daa. When he sysem is done, PIN_RB_RE, PIN_RB_RADD PIN_RB_RDATA are used o access he DPRAM o rerieve he sored maernal feal RR inerval resuls from heir corresponding memory segmen. 3 Resuls Discussion 3. Simulaion Resuls Comparison The resul in Table using Visual C++ shows encouraging resuls wih he es case, where he feal R peak could be deeced up o 98%. Upon compleion of he es case simulaion using VHDL, he sysem performs a read reques o rerieve all he maernal feal resuls. DATA NEWDATA PIN_DATAREQ Daa Inpu Main Conrol Iniializaion XCM IniPar Inipar2 IniSub RR Inerval Validae Subrac Correlaion & Local Maxima PIN_RB_RE PIN_RB_RADD PIN_RB_RDATA Regiser & Memory ECG Signal Daa Templae Templae Parameers Parameers RR Resuls RR Resuls CLK RESET Memory Iniializaion Iniializaion Sae Machine Correl. Loc. Max Templae Updae Coincide Exrac Sae Machine Fig. : Top Level Diagram

5 Table : Visual C++ Tes Case Simulaion Resuls Descripion Value Toal R Peak 64 Deeced 64 (00%) Toal R Peak 235 Deeced Using Firs Threshold 82 (77%) Deeced Using Second Threshold 5 (02%) Coincidence 46 (9%) A sample read operaion for maernal feal R peaks deecion is shown in Fig. 2 wih he iniial 4900 sample daa. The Quarus simulaion resul shows ha he VHDL models are funcioning almos similar o he Visual C++ funcion. The resuls for boh versions are shown in Table 2 3. Comparing he maernal feal RR inerval values (in erms of number of samples beween he inervals), he maernal error is consisenly less han 0.3%, he feal error percenage is wihin 0.5%. All he differences are caused by he rounding effec during compuaion. However, when a feal peak loss happens, an error rae up o 4% migh be occurred, owing o slighly differen search limi implemened in he VHDL. Despie his, he VHDL inerpreaion of he sysem displays grea similariies o he Visual C++ version. Table 2: Visual C++ versus VHDL Resuls for RR Inerval RR Inerval No (Number of Samples) % Diff. VC++ VHDL % % % % % % % % % % % % % % 4 Conclusion The performance achieved for he hear rae measuremens from he AECG shows ha he model can exrac boh maernal feal hear raes uilizing a single-lead configuraion. The single-lead feaure is desirable from he comfor poin of view of he paien especially when he duraion of monioring is long. Table 3: Visual C++ versus VHDL Resuls for RR Inerval RR Inerval No (Number of Samples) % Diff. VC++ VHDL % % % % % % % % % % % % % % % Some improvemens o he R peak deecion capabiliy of he algorihm would be expeced wih enhanced procedures such as he normalizaion of he cross-correlaion oupus variable MECG complex duraion o ake ino accoun for cases when he MHR exceeds 25 BPM. The sensiiviy of he algorihm o moion arifacs muscle noise may also be reduced wih he incorporaion of more rules in is RR inerval validaion schemes. As expeced, a favorable FECG s signal-o-noise raio is a definiive enhancemen in fuure research. An ineresing direcion for fuure research is o compare he oucome of clinical diagnosis based on FHR/MHR deerminaion using our sysem o FHR deerminaion by ulrasound. Anoher area for furher invesigaion could be he case of wins resoluion by AECG. When fully developed, such a sysem will be a useful ool in he assessmen of he feal condiion is relaionship o ha of he moher s. References: [] R. K. Freeman, T. J. Garie M. P. Nageoe, Hear Rae Monioring, Lippinco Williams & Wilkins: Philadelphia, USA, [2] J. A. Crowe, B. R. Hayes-Gill D. K. James, Towards Hear Monioring a Home using he Abdominal Elecrocardiogram, Proc. of IEEE Engineering in Medicine Biology Sociey, Amserdam, 997, Vol., pp

Value =323 (Missing R Peak) Value = 23 (Coincidence) Value = 89 (Coincidence) RR 2 3 4 5 6 7 8 9 0 2 3 4 5 RR 2 3 4 5 6 7 8 9 0 2 3 4 Fig. 2: A Sample Resul of R-peaks Deecion (daa: -4900). [3] A.

Ahmed, M. A. Mohd Ali E. Zahedi, Real-Time Signal Processing for Hear Rae Monioring, IEEE Trans. on Biomedical Engineering, Vol. 50, No. 2, 2003, pp. 258-262. [5] S. Abboud, G. Barkai, S. Mashiach, D.

6 Value =323 (Missing R Peak) Value = 23 (Coincidence) Value = 89 (Coincidence) RR RR Fig. 2: A Sample Resul of R-peaks Deecion (daa: -4900). [3] A. Khamene S. Negahdaripour, A New Mehod for he Exracion of ECG from he Composie Abdominal Signal, IEEE Trans. on Biomedical Engineering, Vol. 47, No. 4, 2000, pp [4] M. I. Ibrahimy, F. Ahmed, M. A. Mohd Ali E. Zahedi, Real-Time Signal Processing for Hear Rae Monioring, IEEE Trans. on Biomedical Engineering, Vol. 50, No. 2, 2003, pp [5] S. Abboud, G. Barkai, S. Mashiach, D. Sadeh, Quanificaion of he Elecrocardiogram Using Averaging Technique, Compuers in Biology Medicine, Vol. 20, No. 3, 990, pp [6] M. A. Mohd. Ali, Developmen of a Porable Hear Rae Recorder, Ph.D. Thesis, Noingham Universiy, UK, 994. [7] J. Pan W. J. Tompkins, A Real-Time QRS Deecion Algorihm, IEEE Trans. on Biomedical Engineering, Vol. 32, No. 3, 985, pp

P. Bruschi: Project guidelines PSM Project guidelines.

P. Bruschi: Project guidelines PSM Project guidelines. Projec guidelines. 1. Rules for he execuion of he projecs Projecs are opional. Their aim is o improve he sudens knowledge of he basic full-cusom design flow. The final score of he exam is no affeced by