Printer Friendly

Classification of Parameters Extracted from Cardiotocographic Signals for Early Detection of Metabolic Acidemia in Newborns.

I. INTRODUCTION

The fetus is normally equipped with physiological defense mechanisms that prevent the long term onset of fetal oxygen deficiency during labor, but in some cases these mechanisms may be weakened, leading to fetal distress. The lack of oxygen that reaches his blood stream through the umbilical cord determines an adaptation of the metabolic processes from normal aerobic metabolism to anaerobic metabolism. The latter has toxic byproducts such as organic acids (lactic acid) that can't be eliminated easily [1]. This may lead to metabolic imbalances (metabolic acidosis) and further in time to neurological dysfunctions, cerebral palsy or encephalopathy [2]. Metabolic acidosis is reflected by the clinical parameters umbilical pH value and base deficit value (BDecf). The method commonly used to determine these values is the fetal scalp blood sampling, an invasive procedure that is time consuming, making it impossible to be used multiple times during labor, and also presents a high risk for the fetus (it requires to take blood samples from the fetus, with possible complications such as hemorrhages, infections or early labor) [3].

In obstetrics, cardiotocography (CTG) is a method of recording the fetal heart rate (FHR) and the uterine contractions (UC) during pregnancy. These signals reflect changes in fetal behavior and are important parameters used in most fetal status assessment techniques. Spectral analysis of FHR and the identification of FHR decelerations coupled with UC is widely used to monitor the autonomic nervous system of the fetus that may cause changes in FHR during oxygen deficiency [4]. FHR monitoring is one of the most used methods that help the physician to diagnose possible abnormalities, and recognize the pathologic conditions during pregnancy stages [5].

The analysis of CTG signals is an area of interest for monitoring maternal and fetal physiological parameters. The literature contains a series of articles that have the main focus of identifying and classifying the CTG signals, in order to determine the state of the fetus at birth. The primary disadvantage of implementing such methods is the inconsistency of the databases used in the evaluation of presented methods. These vary from small local databases to large ones that have signals selected based on clinical and technical criteria. Table I presents a series of noteworthy articles from 2009 to 2015 on the subject of CTG signals analysis and the results obtained. The main purpose of these papers was to classify the signal in two classes: normal or abnormal (acidemic, adverse or pathological). The methods are different in each paper, but the comparison between results is done by means of sensitivity (SE) and specificity (SP) of the method [6].

The best results are obtained by Siira et al. (2012) with SE 89% and SP 80%. The downside of this article is the low number of pathological recordings (15 out of 334). For each article we determined the precision with the geometrical average, Gmed:

Gmed = [square root of SE X SP] (1)

II. DATABASE AND METHODS

We utilized an open access database obtained from Physionet. The CTU-UHB Intrapartum Cardiotocography Database which contains 531 recordings (out of 552; 21 recordings were eliminated based on high signal loss ~80% or lack of clinical annotations) that were collected between 2010-2012 at the University Hospital in Brno (Czech Republic). The recordings consist of fetal heart rate (bpm) and uterine pressure (mmHg) signals from patients with singleton pregnancies, during labor (Figure 1) [12]. These patients didn't have prior problems and the majority of deliveries were normal, only 46 were done through cesarean section. The recordings have an annotation file that specifies the clinical context of the recording, the delivery means, as well as the fetal and neonatal state, as it is presented in Table II.

The FHR signal presents signal loss artifacts that appear due to bad contact between the electrode and the fetal scalp. Fetal heart rate variability (fHRV) analysis depends on the quality of input signal. The FHR measurements may be distorted by fetal or maternal movements, uterine contractions, or misplaced fetal electrode, leading to a corrupted input signal, unusable for automatic computerized analysis [13]. Thus, for each recording we computed the value of the signal loss ratio (SLR) using the formula:

SLR = (SL / TS) X 100, (2)

where SL represents the signal loss and Ts represents the total number of signal samples, respectively. The SLR has an average of 18.6% over the entire database, with a maximum value of 53.5%.

In order to improve results we applied a series of preprocessing steps. The first step consists in smoothing the FHR data with an average over a 30 samples window. The second step identified the signal segments that are not affected by signal loss by automatically removing samples equal to 0. We imposed a minimum length of 2 minutes (480 samples) for each segment (length chosen from the influence of uterine contractions over the FHR signal) [14].

Based on pH and base deficit (BDecf) we defined two classes: normal recordings (pH [greater than or equal to] 7.2 and BDecf [less than or equal to] 8 mmol/L) and abnormal recordings (pH < 7.2 and BDecf > 8 mmol/L). We divided the database into two groups: a control group of 100 normal recordings and a test group that contains the remaining 431 recordings (both normal and abnormal).

The decomposition of the fetal heart rate into frequency components is a useful tool in bio-signal processing field. Using the Fast Fourier Transform, we computed the power spectral density for three non-overlapping energy bands [15]: low frequency LF (0.03-0.07 Hz), mid-frequency MF (0.07-0.13 Hz), and high frequency HF (0.13-1 Hz). In abnormal cases (pH < 7.2 and BDecf > 8 mmol/L) the computed statistical indicators (normalized MF-nMF, normalized HF-nHF and MF/HF ratio) show higher values than in normal situations [16]. We have used the normalized values of the mid and high frequency bands and the mid and high frequency ratio (the sympatho vagal influence of the nervous system over the heart rate) as the main parameters because these are best suited to highlight the heart rate variations of the fetus [17].

Only the last two bands (MF and HF) were selected to perform the signal analysis because these provide the most useful information in regards to fetal physiological activity. We apply the FFT over a 480 samples window with a 240 samples overlap for each recording, only for the signal segments without signal loss [18]. The values obtained for each segment are normalized accordingly with equations (3) and (4). We also determined the medium frequency to high frequency ratio, which was computed using formula (5) [19].

nM[F.sub.i] = MF/(LF+MF+HF) X 100 (3)

nH[F.sub.i] = HF/(LF+MF+HF) X 100 (4)

MFH[F.sub.i] = MF/HF (5)

We then proceeded to compute global values for each recording for the nMF, nHF and MFHF statistical indicators by using a weighted average. For instance,

nMF = [n.summation over (i=1)] [s.sub.i]/N X nM[F.sub.i], (6)

where [s.sub.i] is the length of the analyzed segment without signal loss, N the length of the FHR signal for the analyzed recording, n - the number of segments and nM[F.sub.i] - the normalized medium frequency component value for the analyzed segment "i".

nHF = [n.summation over (i=1)] [s.sub.i]/N X nH[F.sub.i], (7)

where nH[F.sub.i] is the normalized high frequency component value for the analyzed segment "i".

MFHF = [n.summation over (i=1)] [s.sub.i]/N X MFH[F.sub.i], (8)

where MFH[F.sub.i] is the MF to HF ratio for the "i" segment of the recording.

For this analysis we also determined the presence of FHR decelerations influenced by uterine contractions. Decelerations are periodic decreases of FHR associated with uterine contractions. Based on the reactivity and behavior in connection to contraction, decelerations can be categorized as: early, late and prolonged decelerations [20].

Early decelerations occur at the onset of a contraction and don't determine a difference larger than 40 bpm from the baseline. The baseline FHR recovers by the end of contraction, hence early decelerations aren't associated with pathological cases.

Late decelerations are associated with decreased uterine blood flow and are result of: hypoxia, cord compression, excessive uterine activity, maternal hypotension, etc. Late decelerations have the lowest point of FHR value, approximately 15 s after the uterine contraction peak.

Prolonged decelerations are characterized as a reduction of FHR value greater than 30 bpm for a period between 2 and 10 minutes. These are caused by the lack of oxygen delivery to the fetus and are associated with high risk situations and poor fetal outcome. They are caused by maternal hypotension, cord compression or uterine hypertonia [21].

The preprocessing stage for the UC signal requires smoothing with a 30 samples window, followed by the identification of contractions and their limits (beginning and end of contraction). We used a peak finder algorithm implemented in Matlab 2013 to identify the contraction peaks and durations (Figure 3) [22].

The FHR decelerations are identified after we have established the FHR baseline. The first step in detecting the baseline was to use the cubic spline interpolation to eliminate signal loss, followed by smoothing of data with a 30 samples window centered on the analyzed sample.

The second step was to determine the first derivative of the FHR signal in order to identify abrupt changes. The derivative samples that are in areas where the value is higher than 1 bpm/s represent high amplitude of the FHR signal and are eliminated from the original data (Figure 2) [23].

For the third step, the remaining samples that have not been removed in the second step are used to determine the average FHR value according to formula (9). True baseline segments are those longer than 15 seconds (60 samples) with a difference between the average FHR value and the segment values in the interval of [+ or -]10 bpm.

Average_baseline = [summation]Identified baseline samples/Number of samples (9)

The final step is to connect the identified baseline segments by means of linear interpolation.

We applied the rules described for each type of deceleration to establish the number and type of deceleration present on each FHR signal.

The nMF, nHF, and MFHF values are obtained for all the recordings, both from the control group and the test group. We perform a t-test [24] with the p value < 0.05 on the control group in order to establish the consistency of the control group. The results of the test presented in Table III show that the recordings in the control group have frequency domain metrics that were not significantly different from respective mean values.

A second t-test with a p value < 0.05 was performed between the control group and the test group. The test group recordings are classified as normal and abnormal. The final results were confirmed based on pH and BDecf values. The recordings are identified as following: TP - true positive: recordings that were correctly classified as abnormal, TN - true negative: normal records, classified accordingly, FP - false positive: normal records classified as abnormal, FN - false negative: abnormal recordings that were determined as normal.

III. EXPERIMENTAL RESULTS

For the recordings in the control group and the test group (normal and abnormal) we have computed the histogram of values for nMF, nHF, and MFHF as it is represented in figures 4, 5, 6.

The classification of the test group into normal and abnormal recordings by comparison to the frequency domain metrics in the control group is summed up in Table IV. Based on these values we evaluate the sensitivity and specificity of our method.

SE = TP / (TP + FN) (10)

SP = TN / (TN + FP) (11)

Figures 7, 8 and 9 represent 3D scatters of the frequency domain metrics for recordings in the test group based on pH and base deficit. These images indicate a separation of the two types of recordings (normal and abnormal) based on the specified clinical outcomes.

Using the results represented in Table IV, we obtained a sensitivity of 98.38% and a specificity of 67.47%. The low specificity level is due to the need to select for analysis only segments without signal loss. This improves the results of the usual approach of cubic spline interpolation of the lost signal, method which introduces incoherent data in the fHRV spectral analysis, but it still doesn't satisfy the computational need of the method. In order to improve the results we associate this method with the identification of decelerations in FHR signal tracing.

We identify each type of deceleration in every recording of the test group and we obtain the statistical results in Table V.

Figures 10, 11 and 12 present the association between uterine contractions and each type of deceleration. Early decelerations are considered normal physiological events; on the other hand the late and prolonged decelerations are pathological events.

The presence of late and prolonged decelerations of the FHR signal are used as an additional confirmation tool in identifying abnormal recordings (pH < 7.2 and BDecf > 8mmol/L). This condition allows the identification of false positive and false negative recordings, increasing the overall performance of the classification to 98.38% sensitivity and 78.24% specificity, with the class distribution shown in Table VI.

IV. DISCUSSIONS

Due to high values of signal loss, i.e. an average of 18.6% across the entire database with a maximum value of 53.5%, it is necessary to identify a method that allows the improvement of spectral analysis results. The most used method is the cubic spline interpolation of missing signal, but this method modifies the values of frequency domain metrics (nMF, nHF and MFHF). in consequence, we implemented a method that improves these results. This consists of identifying the signal segments without loss and computing the FFT for these segments. The global value for each record is determined by a weighted average, with the weight represented by the ratio of the segment length to the signal length.

We approached this study by taking into consideration two clinical outcomes that are indicative to the fetal status at birth: pH value and base deficit value. Based on these indicators we defined two classes: normal recordings (pH [greater than or equal to] 7.2 and BDecf [less than or equal to] 8 mmol/L) and abnormal recordings (pH < 7.2 and BDecf > 8 mmol/L).

The database was divided into two groups: a control group of 100 normal recordings and a test group with the remaining 431 both normal and abnormal recordings. This allowed us to perform a t-test between the two groups in order to classify the test group recordings into normal and abnormal based on frequency domain metrics. The obtained results point towards a good separation of classes based on these metrics. We initially obtained a sensitivity value of 98.38% and a specificity of 67.47%. To improve the result we introduced a new condition that removes some of the false positive and false negative recordings, which consists in the presence of late and prolonged decelerations associated with uterine contractions in the FHR tracing.

Although the methodology presented in this study is widely used, the overall aim was novel.

V. CONCLUSIONS

Spectral parameters may be used to establish significant differences between normal and pathological fetal outcomes based on pH and base deficit values. Also they can be used to predict the fetal outcome problems that may occur. The overall performance of the classification method is improved when introducing the criteria of late and prolonged decelerations associated with uterine contractions, to 98.38% sensitivity and 78.24% specificity. The presence of early decelerations indicates a normal recording and allows us to identify false positive recordings, and the presence of late and prolonged decelerations indicates an abnormal recording and allows us to remove false negative classifications.

Future work on the identification of fetal acidosis can be done and improved by using data collected from patients that are not in labor, in order to minimize the amount of signal loss induced by fetal and maternal movement and intense uterine contractions. This can be realized by long-term CTG monitoring with a non-invasive device during pregnancy (starting with the 25-28th week of pregnancy).

REFERENCES

[1] A. C. Gjerris, J. Staer-Jensen, J. S. Jorgensen, T. Bergholt, C. Nickelsen,"Umbilical cord blood lactate: A valuable tool in the assessement of fetal blood acidosis," European Journal of Obstretics & Gynecology and Reproductive Biology, vol. 139, Issue 1, pp. 16-20, Jan. 2008. [Online]. Available: http://dx.doi.org/10.1016/j.ejogrb.2007.10.004

[2] E. Soncini, S. Paganelli, C. Vezzani, G. Gargano, G. Battista, "Inatrapartum fetal heart rate monitoring: evaluation of standardized system of interpretation for prediction of metabolic acidosis at delivery and neonatal neurological morbidity," The Journal of Maternal-Fetal & Neonatal Medicine, vol. 27, no. 14, pp. 1465-1469, Sept. 2014. [Online]. Available: http://dx.doi.org/10.3109/14767058.2013.858690

[3] E. Chandraharan, "Fetal scalp blood sampling during labor: is it a useful diagnostic test or a historical test that no longer has a place in modern clinical obstretics?," BJOG: An international Journal of Obstreticts & Gynaecology, vol. 121, Issue 9, pp. 1056-1062, Aug. 2014. [Online]. Available: http://dx.doi.org/10.1111/1471-0528.12614

[4] J.Y.Kwon, I. Y. Park, J.C. Shin, J. Song, R. Tafreshi, J. Lim,"Specific changes in spectral power of fetal heart rate variability related to fetal acidemia during labor: Comparison between preterm and term fetuses," Early Human Development, vol. 88, Issue 4, pp. 203-207, April 2012. [Online]. Available: http://dx.doi.org/10.1016/j.earlhumdev.2011.08.007

[5] M.P. Nageotte, "Featl heart rate monitoring," Seminars in Fetal & Neonatal Medicine, vol. 20, pp. 1-5, Mar. 2015. [Online]. Available: http://dx.doi.org/10.1016/j.siny.2015.02.002

[6] A. Indrayan, "Medical Biostatistics, Third Edition", Chapman & Hall/CRC Press, USA, pp. 280-283, 2012

[7] A. Costa, D. Ayres-de-Campos, F. Costa, C. Santos, J. Bernardes "Prediction of neonatal acidemia by computer analysis of fetal heart rate and ST event signals", American Journal of Obstetrics and Gynecology, vol. 201, pp. 464-452, Nov. 2009. [Online]. Available: http://dx.doi.org/10.1016/j.ajog.2009.04.033

[8] C. Elliott, P. Warrick, E. Graham, E. Hamilton, "Graded classification of fetal heart rate tracings: association with neonatal metabolic acidosis and neurologic morbidity," American Journal of Obstetrics and Gynecology, vol. 202, no. 3, pp. 258.e1-258.e8, Mar. 2010. [Online]. Available: http://dx.doi.org/10.1016/j.ajog.2009.06.026

[9] S. Siira, "Intrapartum hypoxia and power spectral analysis of fetal heart rate variability," Uniprint Suomen Yliopistopaino Oy - Oulu, Finland, pp. 33-42, 2012

[10] J. Spilka V. Chudacek, M. Koucky, M. Huptych, P. Janku, G. Georgoulas, C. Stylios, "Using nonlinear features for fetal heart rate classification," Biomedical Signal Processing and Control, vol. 7, Issue 4, pp. 350-357, July 2012. [Online]. Available: http://dx.doi.org/10.1016/j.bspc.2011.06.008

[11] A. Georgieva, S. J. Payne, M. Moulden, C. W. G. Redman. "Artificial neural networks applied to fetal monitoring in labour,". Neural Computing and Applications, vol. 22, pp. :85-93, Jan. 2013. [Online]. Available: http://dx.doi.org/10.1007/s00521-011-0743-y

[12] V. Chudacek, J. Spilka, M. Bursa, et al., "Open access intrapartum CTG database", BMC Pregnancy and Childbirth, pp. 14:16, Jan. 2014. [Online]. Available: http://dx.doi.org/10.1186/1471-2393-14-16

[13] G.S. Dawes, M. Lobb, M. Moulden, C.W. Redman, T. Wheeler, "Antenatal cardiotocogram quality and interpretation using computers," BJOG: An International Journal of Obstretics & Gynaecology, vol. 99, Issue 10, pp. 791-797, Aug. 2005. [Online]. Available: http://dx.doi.org/10.1111/j.1471-0528.1992.tb14408.x

[14] P.A. Warrick, E.Ff Hamilton, D. Precup, R. Kearney, "Classification of normal and hypoxic fetuses from systems modeling of intrapartum cardiotocography," IEEE Transactions on Biomedical Engineering, vol. 57, Issue 4, pp. 771-779, April 2010. [Online]. Available: http://dx.doi.org/10.1109/TBME.2009.2035818

[15] E.M. Graatsma, "Monitoring of Fetal Heart Rate and Uterine Activity", Ridderprint BV, Amsterdam, Holland, pp. 39-55, 2010

[16] C-Y. Chen, C. Yu, C-C. Chang, C-W. Lin, "Comparison of a Novel Computerized Analysis Program and Visual Interpretation of Cardiotocography," PLoS ONE, vol. 9, Issue 12, Dec. 2014. [Online]. Available: http://dx.doi.org/10.1371/journal.pone.0112296

[17] U. Schneider, E. Schleussner, A. Friedler, S. Jaekel, M. Liehr, J. Haueisen, D. Hoyer, "Fetal heart rate variability reveals defferential dynamics in the intrauterine development of the sympathetic and parasympathetic branches of the autonomic nervous system," Physiologcal Measurements, vol. 30, no. 2, pp. 215-226, Jan. 2009. [Online]. Available: http://dx.doi.org/10.1088/0967-3334/30/2/008

[18] V. Munteanu, D. Tarniceriu, "Estimation theory and optimal filtering," Ed. Technopress, Iasi, Romania, pp. 306-310, 2005

[19] V. Maier, S. G. Pavel, C. D. Maier, I. Birou, "Correct Application of the Discrete Fourier Transform in Harmonics," Advances in Electrical and Computer Engineering, vol. 8, no. 1, pp. 26-30, 2008, doi:10.4316/AECE.2008.01005

[20] M. Jezewski, R. Czabanski, J. Wrobel, K. Horoba, "Analysis of extracted cardiotocographic signal features to improve automated prediction of fetal outcome", Biocybernetics and Biomedical Cardiology, vol. 30, no.4, pp. 29-47, Feb. 2010.

[21] A.G. Cahill, K. A. Roehl, A. O. Odibo, G. A. Macones, "Association and prediction of neonatal acidemia," American Journal of Obstretics and Gynecology, vol. 207, Issue 3, pp. 206.e1-206.e8, Sept. 2012. [Online]. Available: http://dx.doi.org/10.1016/j.ajog.2012.06.046

[22] Y. Hatakeyama, H. Kataoka, N. Nakajima, T. Watabe, Y. Okuhara, "Level evaluation system for cardiotocography," 15th International Symposium on Soft Computing and Intelligence Systems, pp. 265-269, Dec. 2014. [Online]. Available: http://dx.doi.org/10.1109/SCIS-ISIS.2014.7044686

[23] L. Jimenez, R. Gonzalez, M.J. Gaitan, S. Carrasco, C. Vargas, "Computerized algorithm for baseline estimation of fetal heart rate," Computers in Cardiology, vol. 29, pp. 477-480, Sept. 2002. [Online]. Available: http://dx.doi.org/10.1109/CIC.2002.1166813

[24] H. J. Seltman, "Experimental design and analysis," Carnegie Melon University, Chapter 6, pp. 141-161, Nov. 2014

Cristian ROTARIU (1), Hariton COSTIN (1,2), Alexandru PASARICA (3), Dragos NEMESCU (1)

(1) Grigore T. Popa University of Medicine and Pharmacy, Iasi, 700115, Romania

(2) Institute of Computer Science, Romanian Academy, Iasi, 700054, Romania

(3) Gheorghe Asachi Technical University, Iasi, 700506, Romania

cristian.rotariu@bioinginerie.ro

10.4316/AECE.2015.03023
TABLE I. ARTICLES ON THE SUBJECT OF CTG SIGNAL ANALYSIS

Article        Criteria  Method       Classes     SE(%)  SP(%)  Gmed(%)

Costa et al.   pH<7.05   Multiscalar  Normal/     83     83     83
2009 [7]                 entropy      acidemic
Elliot et al.  BDecf>12  PeriCalm     Normal/     57     97     74,36
2010 [8]                 CTG          abnormal
Siira et al.   pH<7.05   ANOVA        Normal/     89     80     84,38
2012 [9]                              acidemic
Spilka et al.  pH<7.15   Support      Normal/
2012 [10]                vector       Pathologic  73     76     74,48
                         machine
Georgieva et   pH<7.1    Neural       Normal/     61     68     64,40
al. 2013 [11]            networks     adverse

TABLE II. CLINICAL ANNOTATIONS OF RECORDINGS

Annotation               Average  Minimum  Maximum

Maternal age               29       18       42
Gestational age (weeks)    40       37       43
pH                          7.23     6.92     7.47
BDecf (mmol/L)              4.43     0.6     23.75
Apgar1/ Apgar5              8/9      3/6     10/10
Weight (g)               3348     1970     4750
Signal length (min)        60       55       95

TABLE III. T-TEST RESULTS FOR THE FREQUENCY DOMAIN METRICS OF THE
RECORDINGS IN THE CONTROL GROUP

HRV                                Confidence intervals for
metric  t statistic  Mean   SD     mean %

                                   90%      27.71 to 32.47
nMF     0.06         30.09  14.3   95%      27.23 to 32.94
                                   99%      26.31 to 33.86
                                   90%      23.15 to 27.22
nHF     0.15         25.19  12.2   95%      22.75 to 27.62
                                   99%      21.97 to 28.41
                                   90%       1.17 to 1.28
MF/HF   0.05          1.23   0.34  95%       1.16 to 1.31
                                   99%       1.14 to 1.32

TABLE IV. TEST GROUP CLASSIFICATION

Record type  Determined records  Reference records

Abnormal     TP          FN       62
              61           1
Normal       TN          FP      369
             249         120
Total        431                 431

TABLE V. NUMBER OF DECELERATIONS IN RECORDINGS OF THE TEST
GROUP

Deceleration type        Min count  Max count  Average

Early decelerations      5          44         6
Late decelerations       0          26         9
Prolonged decelerations  0           7         0.2

TABLE VI. TEST GROUP CLASSIFICATION AFTER INTRODUCING FHR
DECELERATION CONDITION

Record type  Determined records  Reference records

Abnormal     TP          FN       62
              61          1
Normal       TN          FP      369
             289         80
Total        431                 431
COPYRIGHT 2015 Stefan cel Mare University of Suceava
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2015 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Rotariu, Cristian; Costin, Hariton; Pasarica, Alexandru; Nemescu, Dragos
Publication:Advances in Electrical and Computer Engineering
Date:Aug 1, 2015
Words:4083
Previous Article:Characteristics of Overvoltage Protection with Cascade Application of Surge Protective Devices in Low-Voltage AC Power Circuits.
Next Article:Wavelength Converters Placement in Optical Networks Using Bee Colony Optimization.
Topics:

Terms of use | Privacy policy | Copyright © 2020 Farlex, Inc. | Feedback | For webmasters