Value of cervical electrical impedance spectroscopy to predict spontaneous preterm delivery in asymptomatic women: the ECCLIPPx prospective cohort study

Preterm birth (PTB) accounts for two‐thirds of deaths of structurally normal babies and is associated with substantial lifetime healthcare costs. Prevention of PTB remains limited by the modest accuracy of prediction methods, namely transvaginal ultrasound (TVS) cervical length (CL) measurement and quantitative cervicovaginal fetal fibronectin (FFN) estimation. We report the first substantive study detailing the predictive performance of a cervical probe device based on electrical impedance spectroscopy (EIS) for PTB – the EleCtriCaL Impedance Prediction of Preterm birth by spectroscopy of the cervix (ECCLIPPx) study. We aimed to compare the accuracy of cervical EIS‐based prediction of spontaneous PTB with that of prediction using TVS‐CL and FFN in asymptomatic women in the mid‐trimester.


INTRODUCTION
Preterm birth (PTB) affects 15 million babies (6-18% of live births) annually. Early (before 28 weeks' gestation) and late preterm babies suffer neurocognitive deficits [1][2][3] and educational difficulties 4 , respectively, with some developing chronic diseases 5,6 . Although the survival rate following PTB has improved, the incidence of PTB has plateaued or risen 7,8 . Prevention of PTB remains limited by a lack of accurate predictive tests and therapies.
The current mainstay of PTB risk assessment is transvaginal ultrasound (TVS) measurement of cervical length (CL) [9][10][11][12] and determination of cervicovaginal fluid fetal fibronectin (FFN) [13][14][15] . Although both techniques have modest predictive value in high-risk asymptomatic women, their sensitivity is low in low-risk populations [16][17][18] . While findings are variable, most reports suggest that the majority of women with a short cervix (< 15 mm) will not deliver before 32 weeks 11,19 , limiting the utility of CL measurement for the prevention of spontaneous PTB 19,20 , particularly among nulliparous women with a singleton pregnancy 21,22 . The cost-effectiveness of CL and FFN assessment is also unclear 23 .
The effectiveness of interventions based on CLdetermined risk of PTB remains controversial 19 ; progesterone treatment for women with a history of PTB or a short cervix on ultrasound has established limited efficacy for prolonging gestation [24][25][26] and, based on some studies, for preventing adverse neurodevelopmental and other health outcomes [27][28][29] . Given that the syndrome of PTB has multiple etiologies, novel risk assessment modalities may improve stratification of patients and enhance the effectiveness of existing and new preventive interventions 30 .
Assessing the cervical matricellular remodeling changes that precede birth is a logical target for PTB screening. Electrical impedance spectroscopy (EIS) non-invasively quantifies cervical tissue impedance to the passage of a small electrical current. Tissue impedance is influenced by several factors, including cell layering, the intra-and extracellular spaces 31 , cell membrane capacitance 32,33 and tissue hydration 34,35 . There has been increasing application of EIS in screening for cervical 36,37 , oral 38 and breast 39 cancers.
Having demonstrated previously that EIS can detect prelabor cervical remodeling changes [40][41][42][43] , and following the observation in a limited pilot study that mid-trimester cervical impedance in women at risk of PTB seemed to predict delivery before 37 weeks with an accuracy comparable to that of ultrasound-derived CL 44 , we aimed to determine whether cervical EIS can predict spontaneous PTB, as compared to prediction using TVS-CL and/or FFN, in asymptomatic women in the mid-trimester.

Setting
This prospective observational cohort study (the EleC-triCaL Impedance Prediction of Preterm birth by spectroscopy of the cervix (ECCLIPPx) study) was carried out at the Jessop Wing Maternity Unit of the Royal Hallamshire Hospital, Sheffield, UK, a tertiary PTB referral center with approximately 7500 births annually, between January 2014 and August 2016. The study was approved by the Yorkshire & Humber (Sheffield) Committee of the UK National Research Ethics Service (REC Number 13/YH/0167). All study participants gave written informed consent prior to each assessment.

Participants
At 20-22 weeks and 26-28 weeks' gestation, we assessed women deemed to be at high risk of PTB based on a history of spontaneous PTB (defined as one or more previous spontaneous PTBs or mid-trimester pregnancy losses (between 14 and < 37 weeks) 45,46 ). We also assessed at 20-22 weeks women with no history of PTB. Women with a history of an abnormal cervical smear in the previous 3 years, cone biopsy or loop excision of the cervix were excluded, as were women with recent/current cervical infection or vaginal bleeding. Multiple pregnancies and pregnancies with fetal anomaly, cervical cerclage or ongoing progesterone treatment were also excluded. Women who were recruited but who were subsequently treated with vaginal progesterone or cervical cerclage were also excluded from the analysis.

EIS device
We used a Sheffield Mark 5.0 EIS device (Department of Medical Physics and Clinical Engineering, Sheffield Teaching Hospitals Foundation Trust, Sheffield, UK) ( Figure 1), which is an update of previous probes that we employed in earlier work 41,44,[47][48][49] . The Mark 5.0 device incorporates an accelerometer and a pressure sensor for taking measurements at a constant application pressure of 2 Newtons (N) (which was determined from preliminary experiments to be the optimum force for cervical EIS measurements). The probe tip comprises eight electrodes (two tetrapolar configurations) mounted in two rings of 3 mm and 5.5 mm in diameter. The electrodes are fabricated out of 37.5% gold and are 0.6 mm and 1.5 mm in diameter for the inner and outer rings, respectively. The overall tip diameter is 11 mm and, by using a combination of a load cell and a triaxial accelerometer, the device measures and displays the applied force to the Impedance spectroscopy to predict preterm birth 295 Figure 1 Sheffield Mark 5.0 electrical impedance spectroscopy device, showing the device base, which houses the electronics, accelerometer and pressure sensor, and the device probe piece (a), the tetrapolar probe configuration (b) and the complete coupled device ready for use to take measurements (c).
clinician. The measurement process was gated to ensure that the applied force was 2 ± 0.2 N, which corresponds to a pressure of 21 kPa (157 mmHg).
The device measures transfer impedance by applying current at 14 frequencies ranging from 76.3 Hz to 625 kHz in octave increments via an adjacent pair of injecting electrodes, and voltage is measured between the remaining pair of sensing electrodes. Each frequency sweep takes 200 ms and, as a quality-control measure, these are repeated until the standard error (SE) of eight subsequent measurements is below a threshold. This yields a minimum time required to record the full-frequency spectrum of 1.6 s and 3.2 s for both electrode rings. The transfer impedance spectral measurements, together with their variance, is transmitted to the bespoke controlling PC application via Bluetooth technology and stored on a custom central database (ArQ; Scientific Computing, Sheffield Teaching Hospitals, Sheffield, UK).
Cervical EIS measurements are highly repeatable, as demonstrated by a mean (SE) coefficient of variation of impedance measurement of 7.3 (1.1)% 47 in the frequency range 39.5-312.5 kHz, which was determined in our preliminary pilot study 44 to be potentially predictive of PTB. In this electrical frequency range, the mean intraclass correlation coefficient of variation of the Mark 5.0 device was predetermined to be 0.91 (95% CI, 0.74-0.98).
Given that pregnancy is associated with changes in cervical epithelium that include ectropion formation and physiological high-grade metaplasia 50 , we matched the EIS spectra to previously generated templates corresponding to normal squamous and columnar epithelia using a minimum least squares method 51 . As described previously 51 , these templates generated from three-dimensional finite element models of epithelial tissue types 52,53 have been shown to improve discrimination for clinical study outcomes in the cervix 51 and the tongue 38 . Measured EIS spectra were used to construct probability distributions for each tissue type, enabling determination of the relative probability that the measurement corresponded to columnar or squamous epithelium.

Experimental procedure
All measurements were performed by a single operator after structured training on capturing EIS measurements and TVS measurement of CL, employing a standardized experimental protocol. On attendance, each woman had triple high vaginal swabs (sterile Dacron swabs; Deltalab Eurotubo 300 263, Fisher Scientific, Loughborough, UK) taken to obtain cervicovaginal fluid, after the passage of a sterile Cusco's vaginal speculum, with one swab sample used to quantify FFN using the 10Q Rapid FFN analyzer (Hologic, Marlborough, MA, USA) and another sent for bacteriological assessment. The sterile electrical impedance probe was then introduced and gently touched on the anterior lip of the cervix, and a button was pressed to capture data automatically once a steady-state application pressure of 2 N had been attained 48 . Three consecutive measurements were taken in quick succession over about 2 min from each subject. Following removal of the device and the speculum, a transvaginal scan was performed to measure CL, and the shortest of three measurements was recorded. All data obtained were captured automatically in the ArQ database software, which also subsequently captured participant details from clinical records. The operator was blind to the EIS data at the time of capture. Clinicians provided standard clinical care for the participants without knowledge of the EIS results. Ultrasound-indicated cerclage, vaginal progesterone and expectant care were offered for women with a short cervix (< 25 mm), taking into account 296 Anumba et al.
previous pregnancy history, the current gestational age and the woman's preferences.

Study outcomes
The primary prediction outcome of the study was spontaneous delivery before 37 weeks' gestation. The secondary outcome was spontaneous delivery before 32 weeks.

Sample-size estimations
Sample-size estimations were based on data from our pilot study 44 and Jessop Wing birth statistics. We estimated that, in order for EIS to be clinically useful, it should demonstrate reasonable sensitivity and specificity of over 80%. We reasoned that, if cervical EIS predicted PTB with a sensitivity of within ± 10% (i.e. 95% CI from 75 to 95%), we would require 49 women who delivered early to be included in our study population in order to have a sufficient subset of spontaneous PTB cases. Assuming that about 25% of women at high risk of PTB deliver before 37 weeks, as reported in the literature and confirmed in our pilot study, we would need 4 × 49 (i.e. around 200) high-risk women (with a history of spontaneous PTB) to be included. This number would be sufficient for estimating specificity with a reasonable degree of precision, since there would be up to 150 high-risk women who did not deliver prematurely. The precision for the specificity estimate would therefore be within ± 0.06 points, i.e. 95% CI, 79-91%. To assess EIS performance in women with no history of spontaneous PTB, we also recruited 250 women with no risk factors for PTB, a population for which current predictive tests have limited value, yet which accounts for > 50% of PTBs 54 .

Statistical analysis
Data analysis was carried out using SPSS 24 (IBM Corp., Armonk, NY, USA) and MedCalc 14.8.1 (MedCalc Software Ltd, Ostend, Belgium). The data are reported and presented according to the revised STAndards for the Reporting of Diagnostic accuracy studies (STARD) Statement 55 . Descriptive statistics were employed to summarize all quantitative data. The normality of data was assessed using the Kolmogorov-Smirnov test. Categorical outcomes, such as spontaneous PTB before 32 weeks or before 37 weeks, were compared using the χ 2 test (or χ 2 test for trend in the case of ordinal outcomes). Continuous variables were compared using parametric (Student's t-test) or non-parametric (Mann-Whitney U-test) tests, as appropriate.
Multivariate analysis, including multiple linear regression and multiple logistic regression, was employed to exclude subject and test variables that could influence and confound the results for the predictive performance of EIS for PTB. Following analysis of transfer impedance at the 14 studied frequencies, the top five frequencies that showed reduced transfer impedance magnitude in the spontaneous PTB group in this study  were used for further investigation using regression analysis. In addition to these inputs, the full spectra were also matched to templates for normal squamous and columnar tissues, as described previously 51,56 , to give two additional model inputs. Given that a model performs optimally in the dataset employed to generate it, validation in another dataset from a similar population provides a better reflection of the usefulness of the model. We therefore employed these inputs, taken from a random subset (30% of all cases; n = 110) of the spectra measured from both the PTB group and term-delivery group, as a training set to produce a final set of model parameters -the EIS index (consisting of tissue transfer impedance values in the frequency range 19.5-312.5 kHz and template-matched probability estimates for normal squamous and columnar tissues) -which was then used prospectively on the remaining measured spectra (n = 255), to test the performance of the model in distinguishing between women with spontaneous PTB and those with term delivery. The output probability of spontaneous PTB was used to produce graphical receiver-operating-characteristics (ROC) curves of percentage sensitivity against 100 -percentage specificity, in order to determine the predictive accuracy of cervical EIS for spontaneous delivery before 37 weeks, quantified as sensitivity, specificity, positive and negative predictive values, positive and negative likelihood ratios and areas under the curves (AUC). Similar test-accuracy ROC curves were generated for TVS-CL and FFN. For further analysis, the probability of spontaneous PTB based upon the analysis of the cervical EIS spectra were combined with probabilities based upon CL and FFN, measured in the same women, to derive a ROC curve of the combined predictive probability of spontaneous PTB using the three modalities. Kaplan-Meier survival curves of time to delivery were produced to depict the predictive performance of EIS for spontaneous PTB based on an optimal cut-off value of the EIS index. Comparisons of test accuracy were also made for different sub-cohorts of pregnant women.

RESULTS
Of 449 pregnant women recruited, 365 were included in the study. Recruitment details, inclusion and exclusion characteristics and clinical outcomes are summarized in Figure 2. Eighty-four women were excluded from the analysis because of a history of colposcopic cervical treatment, progesterone therapy during the index pregnancy or indicated cervical cerclage for presumed cervical insufficiency. The 365 untreated women included 159 who were deemed to be at high risk for PTB based on a history of spontaneous PTB, while 206 women were deemed to be at low risk for PTB. In this untreated cohort, there were 29 spontaneous PTBs, 14 indicated PTBs and 322 term births (Figure 2). Participant demographic and clinical outcome characteristics are detailed in Table 1.
Smoking prevalence and maternal body mass index were significantly higher in women who delivered before 37 weeks than in those who delivered at term.

Cervical EIS measurements
Consistent with data obtained during an earlier feasibility study 44 , readings obtained from the inner 3-mm tetrapolar electrode ring arrangement showed better separation of findings between the PTB and term-delivery groups and are reported here. In the cohort of women studied, cervical transfer impedance was significantly lower in women who delivered spontaneously preterm than in those who delivered at term when assessed in the frequency range 39.1-312.5 kHz (Table S1). In women at high risk for PTB who were studied at 26-28 weeks, cervical transfer impedance was significantly lower in the subset who delivered preterm when assessed in the frequency range 19.5-312.5 kHz (Table S2).

Model validation data for cervical EIS
The predictive performance of the developed model was similar in the training set and the validation/test set, as summarized in Table 2.

Predictive accuracy of cervical EIS at 20-22 and 26-28 weeks for spontaneous PTB
The combined total cohort of women was used to determine the overall predictive performance for spontaneous PTB of cervical EIS in comparison with TVS-CL and FFN (Figure 3). Cervical EIS, TVS-CL and FFN measured at 20-22 weeks were independently predictive of spontaneous PTB before 37 weeks. Combining all three modalities improved the prediction of spontaneous PTB before 37 weeks, as compared with TVS-CL (P < 0.05) or FFN (P < 0.05) alone (Figure 3a). Furthermore, cervical EIS showed higher predictive accuracy for spontaneous 18

Figure 2
Flowchart showing participant recruitment details, inclusion and exclusion characteristics and preterm birth (PTB) outcomes in low-risk (no previous PTB) and high-risk (previous late miscarriage or spontaneous PTB (14 to < 37 weeks)) groups of asymptomatic women. *Some women had more than one exclusion criterion. GA, gestational age.
PTB than did TVS-CL in women with a short cervix (AUC, 0.83 vs 0.75 for CL < 15 mm (n = 16) and AUC, 0.84 vs 0.53 for CL 15-25 mm (n = 51)). Cervical EIS, TVS-CL and FFN were also predictive of spontaneous delivery before 37 weeks in the subset of high-risk women who were studied again at 26-28 weeks (n = 121) (Figure 3b). In this subgroup, combining all three modalities improved the prediction of spontaneous Data are given as n (%), median (interquartile range) or mean ± SD. *Fourteen women in the preterm-delivery group and 107 in the term-delivery group were assessed at 26-28 weeks. †Data presented only for spontaneous preterm-delivery cases (n = 29). BMI, body mass index; CL, cervical length; FFN, fetal fibronectin; GA, gestational age; PTD, preterm delivery.  PTB before 37 weeks, as compared with TVS-CL (P < 0.05) or FFN (P < 0.05) alone (Figure 3b).
When employing the optimal predictive cervical EIS index cut-off at 20-22 weeks of 0.118 in a binary classification to assess the rate of pregnancy continuation to 42 weeks, Kaplan-Meier survival analysis of time-to-delivery curves showed high predictive performance of cervical EIS (χ 2 , 37.4922; df 1; P < 0.0001) (Figure 4).

Influence of obstetric history on cervical EIS prediction of spontaneous PTB
When taking into account history of spontaneous PTB, the accuracy of prediction of spontaneous delivery before 37 weeks and before 32 weeks was significantly better for cervical EIS than for TVS-CL (P < 0.01 for both) and FFN (P < 0.05 for both), when assessed at 20-22 weeks (Table 3). Furthermore, the predictive accuracy of cervical EIS for spontaneous PTB was not improved further by the incorporation of TVS-CL and FFN.

DISCUSSION
In this study, we detail, for the first time, cervical EIS measurements in the mid-trimester of pregnancy for the prediction of spontaneous PTB (before 37 weeks). In our cohort of asymptomatic pregnant women, we compared EIS to conventional clinical tests -TVS-CL measurement 10,19 and vaginal FFN quantitation 13 . We showed that cervical EIS has a strong predictive potential for subsequent PTB in this untreated cohort, with predictive accuracy comparable with, and in some clinical situations better than, that of TVS-CL and FFN estimation 21 . Larger studies are required to determine whether EIS may prove to be of clinical utility in the prediction of PTB, either as a standalone test or in conjunction with current clinical tests, more so when combined with maternal demographics and a history of spontaneous PTB.
Our findings are in agreement with those of our limited feasibility study of women at high risk of PTB 44 , which similarly showed lower cervical-tissue transfer impedance values in the mid-trimester in those destined to deliver preterm. How cervical remodeling changes influence this observation remains unclear. EIS assesses dielectric tissue properties, which are influenced by cell volume, intra-and extracellular conductivities, cell plasma membrane capacitance 32,57 and the functional state of cellular gap junction proteins 58 . Tissue-transfer impedance also decreases with increasing hydration and edema 43,59,60 . Cervical EIS assesses these properties in the epithelium 51,52 , as well as the subepithelial stroma 49 . Determining the relative contribution of these tissue characteristics to the measured impedance is hampered by the limited information available regarding cervical histologic and microscopic changes during pregnancy. However, it can be speculated that cervical remodeling changes that occur several weeks/months before labor account for the lower tissue-transfer impedance values in women destined to deliver preterm. Recent studies by our group and others have also demonstrated vaginal dysbiosis (characterized by community state types deficient in Lactobacillus species) in women destined to deliver preterm [61][62][63] , which may modulate cervical epithelial and stromal remodeling, leading to PTB 61 .
We have shown that mid-trimester cervical EIS has predictive accuracy for spontaneous PTB that is equal to or better than that of TVS-CL or FFN. Incorporating a history of spontaneous PTB into the risk-assessment model improved the prediction of spontaneous PTB by all assessed modalities, with the highest accuracies being achieved by cervical EIS alone or when combined with TVS-CL and FFN. We have also shown that cervical EIS has higher predictive accuracy for spontaneous PTB than does TVS-CL for women with a shorter cervix in our limited cohorts. Larger studies will determine whether cervical EIS can improve PTB risk assessment in pregnant women with a short cervix (< 15 mm), given that fewer than 1 in 25 of them will deliver before 32 weeks' gestation 11,19 . Such studies will also clarify the predictive performance of cervical EIS for PTB in nulliparous and low-risk women.
In this initial report, we detail the predictive performance of cervical EIS solely in pregnant women receiving no treatment intervention for PTB in order to reduce potential confounding. A larger study will be required to determine the value of cervical EIS in directing or assessing such interventions and in women who have had cervical surgery, those who have a congenital uterine malformation 64 and in multiple pregnancies, given the dearth of adequate PTB risk-assessment approaches for these groups [65][66][67][68] .
The predictive performance of TVS-CL and FFN for PTB before 32 or 37 weeks' gestation in asymptomatic women studied in the mid-trimester in our series is in agreement with that in other reports for unselected pregnant women as well as for those with a history of PTB 16,17,21,22,69 . Furthermore, our limited data also highlight that combining EIS with TVS-CL and FFN for risk assessment in asymptomatic nulliparous women may improve the prediction of PTB, and warrants further study in this group of women, who currently have limited options for PTB risk assessment. If these studies are coupled with effective interventions and demonstrate benefit, they may enable the generation of better predictive algorithms and decision-support tools for the management of PTB 70,71 .
If our observations are confirmed by larger studies, clinical adoption of cervical EIS as a PTB risk-assessment tool may confer several advantages. It would provide a hand-held point-of-care test that would require limited additional training of frontline care practitioners in the maternity healthcare setting. Regardless of the varied etiologies of the PTB syndrome, the EIS device assesses cervical remodeling, which is the final common path to the onset of preterm labor. It could potentially be employed for risk assessment for the majority of pregnant women. It can be employed in obese women subject to possible technical difficulties of visualizing the cervix during a pelvic examination. Minimal training is required, and technical proficiency may be achieved after about 10 measurements by a medically trained person able to perform a pelvic examination. An operating manual and/or training video would suffice to deliver such training. As it is distinct from existing risk-assessment approaches, EIS could prove complementary to them. This study has several limitations. The sample size was limited and precluded assessment of women who had had cervical loop excisional treatment, cervical cerclage or progesterone, as well as those with a multiple pregnancy. The study population was mainly Caucasian, and further studies of women of other ethnicities will be required to demonstrate generalizability of the findings. Since a single operator obtaining all study measurements ensured better reliability of the data, interobserver performance of the measurements was not assessed. However, we have reported previously high inter-and intraobserver repeatability and reproducibility of cervical EIS 47 . The limited sample size also precluded assessing cervical EIS prediction of secondary outcomes such as neonatal morbidity. Larger, multicenter studies will be required to determine the potential value of cervical EIS for routine unselected screening during pregnancy, the effects of treatments on EIS measurements and the potential incorporation of EIS into assessment and treatment algorithms, as well as health economic and value-of-information analyses.
In conclusion, subject to confirmation in larger studies, our observations suggest potential clinical utility of cervical EIS for PTB risk assessment, either as a standalone test or in conjunction with existing modalities.