Ultrasound Obstet Gynecol 2016; 47: 210–216 Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/uog.14894
Development and validation of a tool incorporating quantitative fetal fibronectin to predict spontaneous preterm birth in symptomatic women K. KUHRT, N. HEZELGRAVE, C. FOSTER, P. T. SEED and A. H. SHENNAN Woman’s Health Academic Centre, King’s College London, London, UK
K E Y W O R D S: predictive model; premature labor; quantitative fetal fibronectin; symptomatic women
ABSTRACT Objective To develop a reliable and validated tool for prediction of spontaneous preterm birth (sPTB) in symptomatic women that incorporates quantitative measurements of fetal fibronectin (qfFN) and other relevant risk factors. Methods Data were analyzed that had been collected prospectively from 382 women who presented at an emergency assessment unit between 22 + 0 and 35 + 6 weeks’ gestation with symptoms of preterm labor. Clinicians were blinded to qfFN although they were aware of qualitative fFN results. Parametric survival models for sPTB, with time-updated covariates, were developed for combinations of predictors and the best was selected using the Akaike and Bayesian information criteria. The model was developed on the first 190 consecutive women and validated on the subsequent 192. The estimated probability of delivery before 30, 34 or 37 weeks’ gestation and within 2 or 4 weeks of testing was calculated for each patient and was compared to actual event rates. Predictive statistics were calculated to compare training and validation sets. Results The final model that was selected used qfFN and previous sPTB/preterm prelabor rupture of membranes (PPROM) as predictors. Predictive statistics were similar for training and validation sets and there was good agreement between expected and observed sPTB for all outcomes. Areas under the receiver–operating characteristics curves ranged from 0.77 to 0.88, indicating accurate prediction across all five delivery outcomes. Conclusions sPTB in symptomatic women can be predicted accurately using a model combining qfFN and previous sPTB/PPROM. Clinicians can use this model, which has been incorporated into an App (QUiPP),
to determine accurately a woman’s risk of sPTB and potentially tailor management decisions appropriately. Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
INTRODUCTION Every year 15 million babies are born preterm (< 37 weeks’ gestation)1 worldwide, and 1.1 million die of associated complications2 . Survivors face significant risk of morbidity. Preterm labor is now one of the most costly indications for obstetric admission3 , but most who present with symptoms of preterm labor do not go on to deliver early4 . Due to the limited universal predictive tools available to identify those at highest risk, these women may receive unnecessary interventions, including tocolysis, steroids and admission to hospital, which are expensive and may be associated with detrimental side effects. Accurate prediction of preterm birth in symptomatic women would allow timely and appropriate intervention, potentially reducing neonatal morbidity and mortality. Fetal fibronectin (fFN) has proven a useful predictor of preterm birth in both symptomatic and high-risk asymptomatic women5,6 . fFN is an adhesive glycoprotein found in the amniotic fluid, placental tissue and extracellular component of the decidua basalis adjacent to the placental intervillous space. It is released after mechanical or inflammatory-mediated damage to the placenta or membranes before birth7 . While traditionally a binary test is used for measurement of fFN, which gives a positive or negative result based on a threshold concentration of 50 ng/mL, a novel bedside quantitative fetal fibronectin (qfFN) test has been developed that adds value over the former qualitative measurement. In our previous prospective study of 300 women at 22–35 weeks’ gestation5 , the positive predictive value
Correspondence to: Prof. A. Shennan, Division of Women’s Health, 10th Floor North Wing, St Thomas’ Hospital, London SE1 7EH, UK (e-mail: [email protected]) Accepted: 5 May 2015
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
ORIGINAL PAPER
Predicting preterm birth using quantitative fFN (PPV) for spontaneous preterm birth (sPTB; < 34 weeks’ gestation) improved with increasing fFN thresholds, from 19% to 32%, 61% and 75% at respective fFN thresholds of 10, 50, 200 and 500 ng/mL. In this way, we demonstrated that qfFN provides additional thresholds to define risk compared to the conventional qualitative test. This facilitates improved discrimination of risk and subsequent targeted management of sPTB in symptomatic women. We aimed to produce a reliable and validated tool for determining the risk of sPTB in symptomatic women attending an emergency assessment unit, using a combination of risk factors and qfFN measurement; not limited by fixed thresholds, the tool should be capable of estimating the individual probability of preterm delivery at any given gestational age or within any given number of weeks after testing. The algorithm has been incorporated into an App (QUiPP) for widespread use.
METHODS This was a prospective observational secondary analysis of a population of women enrolled in the ongoing Evaluation of Fetal Fibronectin with a Quantitative Instrument for the Prediction of Preterm Birth (EQUIPP) study involving 382 consecutive women with a singleton pregnancy who presented to an emergency assessment unit between 22 + 0 and 35 + 6 weeks’ gestation with symptoms of threatened preterm labor, as assessed by the attending clinician. Women with a blood-stained swab or sexual intercourse within the last 24 h were excluded from the study due to known interference with fFN measurement. Women with multiple pregnancy or those with insufficient or absent qfFN sample, incomplete outcome data or offspring from the current pregnancy with a major congenital abnormality were also excluded. Participants were recruited from five hospitals in the UK between October 2010 and February 2014. The study was approved by South East London Research Ethics Committee, and the local research ethics committees of all participating centers. Written informed consent was obtained from all participants. The fFN sample was obtained by an obstetrician during a sterile speculum examination in which a polyester swab was inserted into the posterior fornix of the vagina to collect cervicovaginal secretions. One aliquot (200 μL) of the sample was analyzed with the conventional fFN TLiIQ analyzer (Hologic, Marlborough, MA, USA) and a second aliquot (200 μL) of the same sample using the new quantitative Rapid fFN 10Q analyzer (Hologic). Clinicians were made aware of the qualitative TLiIQ result (positive/negative), but both patient and clinician remained blinded to 10Q results until after delivery. Participants’ demographic information, risk factors for preterm birth and obstetric and gynecological history were entered into a secure online database (www.medscinet.net/PTBstudies). Women with a positive TLiIQ fFN result were managed as per unit protocols that included administration of steroids, tocolysis and bed rest.
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
211
Any model performs optimally on the dataset that has been used to generate that model. In order to provide a true assessment of the efficacy of the model, the study population was split into a training set, consisting of 190 women with 208 visits, that was used to develop the model, and a validation set, which included 192 women with 221 visits. A predictive model was run using the validation set to assess the performance of the algorithm.
Statistical analysis Data sets Statistical analysis was performed with Stata software (version 11.2; StataCorp LP, College Station, TX, USA). The data were randomly split 1:1 into training and validation sets, each woman providing a fFN sample on one or more occasion. All test results from all participants were included in the analysis. Model generation Survival analysis with time-updated covariates for preterm birth was used to identify the principal predictors. Women were considered at risk of an event from the time of the visit to the earliest of next test, delivery or 37 weeks’ gestation. Deliveries after 37 weeks or iatrogenic preterm deliveries were regarded as censored (n = 11). fFN, gestational age at test, previous preterm delivery/PPROM and details of symptoms, i.e. abdominal pain or threatened preterm labor (more than one contraction every 10 min for at least 30 min), were considered as possible predictors. We performed an initial analysis in which other predictive variables, including body mass index, ethnicity, smoking, previous cervical surgery and previous late miscarriage, were excluded as not significant. Six parametric survival models were compared for each combination of predictors: exponential, gamma, Gompertz, log-logistic, log-normal and Weibull. The models described the probability of early delivery throughout pregnancy, as it changed according to gestational age between test and delivery. The best survival function was determined by having the lowest values of the Akaike and Bayesian information criteria (AIC and BIC)8,9 . Predictive variables were then investigated for non-linearity using fractional polynomials10 and a series of qfFN cutpoints at standard values: 10, 20, 50, 100, 200 and 500 ng/mL. Model validation To confirm test performance, the Hosmer–Lemeshow test was used to calculate actual and expected event rates by decile of risk for delivery before 30, 34 or 37 weeks’ gestation and within 2 or 4 weeks of testing11 . Receiver–operating characteristics (ROC) curves were drawn and areas under the curve (AUC) calculated. Predictive statistics of the algorithm were calculated, using a probability above 10% as indicating a positive
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Kuhrt et al.
212 Women presenting at emergency assessment unit between 22 + 0 and 35 + 6 weeks with symptoms of threatened preterm labor (n = 504 women, 592 visits)
Excluded (n = 122 women, 163 visits): Multiple pregnancy (n = 25 women, 53 visits) Congenital abnormality (n = 18 women, 22 visits) Sexual intercourse < 24 h (n = 8 women, 9 visits) Blood-stained swab (n = 12 women, 14 visits) Insufficient or absent qfFN measurement (n = 24 women, 28 visits) Incomplete outcome data (n = 35 women, 37 visits)
Final study population after exclusion (n = 382 women, 429 visits)
Training set (n = 190 women, 208 visits)
Validation set (n = 192 women, 221 visits)
Figure 1 Flow diagram of women with threatened preterm labor illustrating the number of women involved in the study after exclusions were made according to the defined criteria. qfFN, quantitative fetal fibronectin.
test in order to compare training and validation sets. This value was chosen as being practically useful for separating low-risk women (probability of delivery around 5%) from high-risk women (probability around 20%) for the purpose of clinical management.
RESULTS After exclusions, as illustrated in Figure 1, the final population of symptomatic women studied consisted of 382 women, and 429 fFN measurements were recorded. This final group was divided into a training set, consisting of 190 women with 208 visits, and a validation set of 192 women and 221 visits. Of the study population, 73% had only a single visit to the unit, at a mean gestational age of 29 weeks; those who had two or more visits typically visited earlier in pregnancy, at a mean gestational age of 25 weeks. The demographic characteristics of the validation and trainings sets were comparable (Table 1) as was the proportion of women in each fFN concentration category (Table 2). In the total study population, 29 (8%) women received tocolysis as per management protocols, two (0.5%) received progesterone and four (1%) received cervical cerclage. The rate of sPTB was 13%, 6%, 3%, 3% and 7% for delivery at < 37, < 34 and < 30 weeks’ gestation and within 2 and 4 weeks of testing, respectively, in the training set and 13%, 7%, 5%, 5% and 10%, respectively, in the validation set. These values are comparable to rates reported in the literature1,2,5 .
Model generation (using the training set only) Of the predictors considered, only qfFN and previous sPTB/PPROM were significant in the stepwise model at P < 0.05 (Table 3). The best parametric survival model
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
Table 1 Demographic characteristics of women with singleton pregnancy presenting with threatened preterm labor in training and validation sets Demographic characteristic Age (years) Ethnicity White Black Asian Other BMI* < 20 kg/m2 20–24.9 kg/m2 25–29.9 kg/m2 ≥ 30 kg/m2 Smoker*† Never Current Stopped Before pregnancy During pregnancy Risk factors Previous sPTB Previous PPROM Previous late miscarriage (16 + 0 to 23 + 6 weeks) Previous cervical surgery
Training set (n = 190)
Validation set (n = 192)
All (n = 382)
30
30
30
92 (48) 68 (36) 11 (6) 19 (10)
92 (48) 65 (34) 20 (10) 15 (8)
184 (48) 133 (35) 31 (8) 34 (9)
27 (14) 81 (43) 41 (22) 41 (22)
32 (17) 73 (38) 48 (25) 37 (19)
59 (16) 154 (41) 89 (23) 78 (20)
138 (73) 23 (12)
148 (78) 23 (12)
286 (76) 46 (12)
19 (10) 9 (5)
15 (8) 4 (2)
34 (9) 13 (3)
36 (19) 13 (7) 14 (7)
28 (15) 9 (5) 14 (7)
64 (17) 22 (6) 28 (7)
14 (7)
12 (6)
26 (7)
Data are given as mean or n (%). There was no significant difference in cervical length between training and validation sets. *Two values missing from validation set. †One value missing from training set. BMI, body mass index; PPROM, preterm prelabor rupture of membranes; sPTB, spontaneous preterm birth.
(determined by lowest AIC and BIC) was a log-normal survival function with terms for qfFN and previous sPTB/PPROM. The best fractional polynomial model was linear for fFN concentration, outperforming those using
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Predicting preterm birth using quantitative fFN
213
Table 2 Fetal fibronectin concentration (FN) for training and validation sets comprising women with singleton pregnancy presenting with threatened preterm labor
Fetal FN
Training set (n = 190)
Validation set (n = 192)
All (n = 382)
114 (60) 20 (11) 21 (11) 10 (5) 10 (5) 15 (8)
113 (59) 19 (9) 15 (8) 9 (5) 18 (9) 18 (9)
227 (59) 39 (10) 36 (9) 19 (5) 28 (7) 33 (9)
< 10 ng/mL 10–19.9 ng/mL 20–49.9 ng/mL 50–99.9 ng/mL 100–199.9 ng/mL ≥ 200 ng/mL
Data are given as n (%). There was no significant difference in fetal fibronectin concentration between training and validation sets. Table 3 Variables tested for inclusion in the model for prediction of spontaneous preterm birth Variable qfFN Previous sPTB/PPROM Gestational age at test Abdominal pain Threatened preterm labor*
Parameter estimate
P
−0.0015 (−0.0024 to −0.0006) −0.18 (−0.35 to −0.013) 0.00038 (−0.0020 to 0.0028) −0.25 (−0.70 to 0.19) −0.16 (−0.64 to 0.31)
0.0002 0.035 0.753 0.264 0.496
Values in parentheses are 95% CIs. *More than one contraction every 10 min for at least 30 min. PPROM, previous preterm prelabor rupture of membranes; qfFN, quantitative fetal fibronectin; sPTB, spontaneous preterm birth.
cutpoints. These findings were confirmed by improvement in AIC and BIC. Therefore the final model included linear fFN and previous sPTB/PPROM (Appendix S1).
Model validation Table 4 summarizes predictive statistics calculated using an estimated probability of delivery > 10% as indicating a positive test, to allow comparison of training and validation sets for prediction of delivery at five clinically important points in time: < 37, < 34 and < 30 weeks’ gestation and within 2 or 4 weeks of qfFN testing. Predictive statistics, including sensitivity, specificity, positive (LR+) and negative (LR–) likelihood ratios were similar for training and validation sets across all five time points considered. The LR+ and LR– values were all much greater than or less than 1, respectively, which indicates that positive test results were strongly associated with the occurrence of sPTB and negative test results with its absence12 . On comparison of observed and expected sPTB rates (Table 5), test performance was generally good. No significant differences were found for women with a probability of early delivery > 10%. For women with a probability ≤ 10%, the observed rate was always < 10%, as required; but it was significantly higher than expected for delivery < 30 weeks and for delivery within 4 weeks.
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
Significance disappears after allowing for multiple testing. However, extremely low probabilities may need to be interpreted with caution. The ROC curves in Figure 2 represent overall test performance regardless of fFN threshold. The AUC values (Table 4), ranging from 0.77–0.88 in the validation set, indicate that the model provides useful information for prediction of sPTB across all five delivery time points investigated.
DISCUSSION We have created an accurate prediction model incorporating qfFN and previous sPTB/PPROM that can be used for the prediction of sPTB in symptomatic women. The model remains accurate when tested on a validation set. Its reliable performance is demonstrated further by comparing expected and observed sPTB rates which gave P-values > 0.05 consistently, indicating that any differences are not significant and are therefore unlikely to be clinically important. This is supportive of previous research that showed that quantification of fFN provides additional information on the risk of sPTB in symptomatic5 and asymptomatic high-risk13 women when compared with the original qualitative fFN test. Although overall sensitivity and specificity are similar to those in the literature, the advantage of the algorithm is that the individual risk for each woman can be determined instead of calculation of a summary value. A previous study including 725 singleton pregnancies demonstrated similar negative predictive values (NPVs) but much lower PPVs; Peaceman et al.14 reported a PPV of 16% for prediction of delivery < 34 weeks’ gestation compared to 33.3% in our study. This suggests that, using our model, clinicians can be more confident in the validity of a positive test result, and can accurately target interventions while avoiding treatment in women who do not need it. The ROC curves (Figure 2) illustrate overall prediction of sPTB regardless of qfFN threshold, which is highly relevant given the use of qfFN in this model, and will provide optimal prediction throughout the range of values on which a clinician can act. AUCs ranged from 0.77 to 0.88, showing an improvement on those reported in previously published literature; Honest et al. performed a systematic review of studies using qfFN at a threshold of 50 ng/mL. Forty studies involving 26 876 women were included and summary ROC areas ranged from 0.71 to 0.77 for prediction of delivery at < 37 and < 34 weeks’ gestation, respectively15 . A major strength of this study is that data were derived from a prospective dataset in which clinicians and patients were blinded to the qfFN result. Furthermore, the parametric method employed to generate the model, combined with appropriate censoring, makes full use of the available data without introducing biases. However, a potential limitation is the small number of women with sPTB at some gestational time points; a larger study could provide the opportunity to further validate our findings.
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Kuhrt et al.
214
Table 4 Predictive statistics for delivery at < 30, < 34 and < 37 weeks’ gestation and delivery within 2 or 4 weeks from fetal fibronectin test in training and validation datasets GA at delivery Parameter Sensitivity (%) Training Validation Specificity (%) Training Validation AUC Training Validation LR+ Training Validation LR– Training Validation PPV (%) Training Validation NPV (%) Training Validation
Delivery within:
< 30 weeks
< 34 weeks
< 37 weeks
2 weeks
4 weeks
100 (47.8–100) 50.0 (18.7–81.3)
66.7 (34.9–90.1) 78.6 (49.2–95.3)
66.7 (44.7–84.4) 72.0 (50.6–87.9)
44.4 (13.7–78.8) 50.0 (18.7–81.3)
68.4 (43.4–87.4) 47.4 (24.4–71.1)
97.8 (94.6–99.4) 96.2 (92.2–98.4)
90.4 (85.1–94.3) 87.6 (81.9–92.1)
78.9 (71.9–84.9) 77.2 (70.1–83.4)
93.4 (88.7–96.5) 94.0 (89.4–96.9)
90.6 (85.3–94.6) 90.2 (84.7–94.2)
1.0 (1.0–1.0) 0.88 (0.74–1.0)
0.76 (0.63–0.89) 0.83 (0.67–0.98)
0.76 (0.63–0.89) 0.77 (0.65–0.90)
0.70 (0.43–0.98) 0.77 (0.55–0.99)
0.80 (0.65–0.95) 0.78 (0.66–0.91)
46.3 (17.5–122.0) 13.0 (5.0–33.8)
7.0 (3.8–12.8) 6.4 (3.9–10.3)
3.2 (2.1–4.8) 3.2 (2.2–4.6)
6.7 (2.7–16.7) 8.3 (3.6–19.2)
7.3 (4.2–12.8) 4.8 (2.5–9.3)
0.09 (0.01–1.21)* 0.5 (0.3–1.0)
0.4 (0.2–0.8) 0.2 (0.1–0.7)
0.4 (0.2–0.8) 0.4 (0.2–0.7)
0.6 (0.3–1.1) 0.5 (0.3–1.0)
0.4 (0.2–0.7) 0.6 (0.4–0.9)
55.5 (21.2–86.3) 41.7 (15.2–72.3)
32.0 (14.9–53.5) 33.3 (18.0–51.8)
31.4 (19.1–45.9) 32.1 (20.3–46.0)
25.0 (7.3–52.4) 31.3 (11.0–58.7)
44.8 (26.4–64.3) 34.6 (17.2–55.7)
100 (98.0–100) 97.2 (93.6–99.1)
97.6 (93.9–99.3) 98.1 (94.6–99.6)
94.2 (89.0–97.5) 94.9 (89.7–97.9)
97.1 (93.4–99.1) 97.2 (93.5–99.1)
96.3 (92.1–98.6) 94.0 (89.2–97.1)
For each gestational time point, an individual probability of early delivery > 10% was treated as a positive test result. Values in parentheses are 95% CIs. *Calculated using substitution formula: 0.5 added to allow for zero values. AUC, area under receiver–operating characteristics curve; GA, gestational age; LR–, negative likelihood ratio; LR+, positive likelihood ratio; NPV, negative predictive value; PPV, positive predictive value. Table 5 Expected and observed rates of spontaneous preterm birth at < 30, < 34 and < 37 weeks’ gestation and within 2 or 4 weeks from fetal fibronectin testing in the validation set (n = 192)
Time point of delivery < 30 weeks P ≤ 10 P > 10 < 34 weeks P ≤ 10 P > 10 < 37 weeks P ≤ 10 P > 10 Within 2 weeks P ≤ 10 P > 10 Within 4 weeks P ≤ 10 P > 10
Event rate (n/N)
Expected rate* (%)
Observed rate (95% CI) (%)
P†
5/180 5/12
0.79 54
2.8 (0.9–6.4) 42 (15–72)
0.02 0.40
3/159 11/33
2.1 40
1.9 (0.4–5.4) 33 (18–52)
1.00 0.49
7/136 18/56
4.5 39
5.2 (2.1–10.3) 32 (20.3–46.1)
0.68 0.34
5/176 5/16
1.3 27
2.8 (0.9–6.5) 31 (11.0–58.7)
0.08 0.78
10/166 9/26
2.5 36
6 (2.9–10.8) 35 (17.2–55.7)
0.01 1.00
For each gestational time point, an individual probability of early delivery ≤ 10% (P ≤ 10) or > 10% (P > 10) was treated as a negative or positive test result, respectively. *Calculated using algorithm. †Comparison of observed and expected rates.
The low rate of tocolysis in the study population likely reflects the practice applied at the units from which patients were recruited, as it is generally used only for in-utero transfer. The small numbers involved are unlikely to impact significantly on sPTB prediction. However, to
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
clarify, 8% of the total study population received tocolysis as per management protocols, which has been shown to prolong pregnancy for 48 h to 7 days16 and 0.5% and 1% received progesterone or cerclage, respectively. Evaluation of the predictive value of fFN independent of decisions regarding these interventions is difficult in current clinical practice, and our prediction model takes this management into account. Reactive therapies such as magnesium sulphate and steroids can still be targeted, but further investigation is required to ascertain how the model predicts without tocolysis, progesterone and cerclage, i.e. clinical evaluation of the model in practice is needed. It is not common practice in the UK to use cervical length as a predictor in the acute setting for symptomatic threatened preterm birth, and we do not have systematic data to include in the algorithm. Future work should incorporate both cervical length and qfFN to obtain a potentially improved model. In the meantime, it remains reasonable to use digital examination and cervical change to ascertain risk of delivery, to confirm true labor and to rule out the rare false-negative fFN test result when clinical suspicion dictates. This subjective assessment has not undergone formal evaluation as a predictive tool and is likely to be superseded in the future by cervical length scanning, particularly in early labor. In practice, the high NPV demonstrated for all gestational time points investigated will enable clinicians confidently to discharge women deemed to be at low risk by the model, without administering costly interventions
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Predicting preterm birth using quantitative fFN
215
0.75
0.75
0.75
0.50
0.25
0
Sensitivity
(c) 1.00
Sensitivity
(b) 1.00
Sensitivity
(a) 1.00
0.50
0.25
0.25
0.25
0.50
0.75
0
1.00
0.25
(e) 1.00
0.75
0.75 Sensitivity
(d) 1.00
Sensitivity
0.50
0.75
1.00
0.50
0.50
0.75
1.00
0.50
0
0.25
1 – Specificity
0.75
1.00
1 – Specificity
associated with potential side effects. Furthermore, the strong overall prediction illustrated by the ROC curves will enable management decisions regarding in utero transfer, tocolytic therapy, antenatal corticosteroids and magnesium sulphate to be tailored appropriately to suit each woman, based on her risk. There is already evidence to show that qualitative fFN sampling in symptomatic women reduces the number of admissions, length of stay and administration of tocolysis. One study demonstrated a 40% reduction in hospital admissions for preterm labor and cost savings of approximately US$486 000, with no negative impact on neonatal outcome17 . With the added value from qfFN testing, admissions and unnecessary interventions could be reduced and costs cut even further. More work is needed to ascertain whether interventions improve outcome when targeted to women identified by the algorithm, but some interventions may be withheld in women deemed to be low risk by the algorithm. The addition of cervical length measurement may further improve prediction and this should be investigated in future studies; we were unable to incorporate it into the current algorithm as we do not have those data. However, another algorithm has been developed for asymptomatic women incorporating qfFN and cervical length18 . Together, these data form the basis of a model for prediction of sPTB in asymptomatic high-risk and symptomatic women for whom the risk of delivery at five clinically important points in time can be calculated accurately based on the qfFN result and previous history of sPTB/PPROM. The algorithm is available through an App (QUiPP) to make it easily accessible to clinicians for use at the bedside and in clinic.
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
0.50
0.25
0.50
0.75
1.00
1 – Specificity
0.25
0.25
0.25
0
1 – Specificity
1 – Specificity
0
0.50
Figure 2 Receiver operating–characteristics (ROC) curves showing overall prediction for delivery < 30 weeks (a), < 34 weeks (b) and < 37 weeks (c), and within 2 weeks (d) or 4 weeks (e) of testing in the validation set, for a model including linear fetal fibronectin and previous spontaneous preterm birth or preterm prelabor rupture of membranes. Areas under ROC curve: (a) 0.8816, (b) 0.8262, (c) 0.7728, (d) 0.7676 and (e) 0.7837.
ACKNOWLEDGMENTS We thank the women who participated in this study, Judy Filmer and the team who assisted with patient recruitment and sample processing from the Preterm Surveillance Clinic at the Division of Women’s Health, King’s College London. This research was supported by Tommy’s Baby Charity (charity number 1060508)
DISCLOSURES Minority financial and equipment assistance was provided by Hologic USA (Marlborough, MA, USA). A.H.S. has received financial assistance from Hologic USA for providing educational talks on preterm birth.
REFERENCES 1. Blencowe H, Cousens S, Oestergaard MZ, Chou D, Moller AB, Narwal R, Adler A, Vera Garcia C, Rohde S, Say L, Lawn JE. National, regional, and worldwide estimates of preterm birth rates in the year 2010 with time trends since 1990 for selected countries: a systematic analysis and implications. Lancet 2012; 379: 2162–2172. 2. March of Dimes, PMNCH, Save the Children, WHO. Born Too Soon: The global action report on preterm birth. Howson CP, Kinney MV, Lawn JE (eds). World Health Organization: Geneva, 2012. http://www.who.int [accessed 29 February 2014]. 3. Russell RB, Green NS, Steiner CA, Meikle S, Howse JL, Poschman K, Dias T, Potetz L, Davidoff MJ, Damus K, Petrini JR. Cost of hospitalization for preterm and low birth weight infants in the United States. Pediatrics 2007; 120: e1–9. 4. Guinn DA, Goepfert AR, Owen J, Brumfield C, Hauth JC. Management options in women with preterm uterine contractions: a randomized clinical trial. Am J Obstet Gynecol 1997; 177: 814–818. 5. Abbott DS, Radford SK, Seed PT, Tribe RM, Shennan AH. Evaluation of quantitative fetal fibronectin test for spontaneous preterm birth in symptomatic women. Am J Obstet Gynecol 2012; 208: 122.e1–122.e6. 6. Bolt LA, Chandiramani M, de Greeff A, Seed PT, Kurtzman J, Shennan AH. The value of combined cervical length measurement and fetal fibronectin testing to predict spontaneous preterm birth in asymptomatic high risk women. J Matern Fetal Neonatal Med 2011; 24: 928–932.
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Kuhrt et al.
216 7. Genc MR, Ford CE. The clinical use of inflammatory markers during pregnancy. Curr Opin Obstet Gynecol 2010; 22: 116–121. 8. Akaike H. Information theory and the extension of the maximum likelihood principle. In Second International Symposium on Information Theory. Petrov VN, Csaki F (eds), Akailseoniai-Kiudo: Budapest, 1973; 267–281. 9. Schwarz G. Estimating the dimension of a model. Ann Stat 1978; 6: 461–464. 10. Sauerbrei W & Royston P. Building multivariable prognostic and diagnostic models: Transformation of the predictors by using fractional polynomials. J R Stat Soc Ser A Stat Soc 1999; 162: 71–94. 11. Hosmer DW Jr, Lemeshow S, Klar J. Goodness-of-fit testing for the logistic regression model when the estimated probabilities are small. Biom J 1988; 30: 911–924. 12. Deeks JJ, Altman DG. Diagnostic tests 4: likelihood ratios. BMJ 2004; 329: 168–169. 13. Kurtzman J, Chandaramani M, Briley A, Poston L, Das A, Shennan AH. Quantitative fetal fibronectin in asymptomatic high risk patients and the spectrum of risk for recurrent preterm delivery. Am J Obstet Gynecol 2009; 200: 263.e1–263.e6.
14. Peaceman AM, Andrews WW, Thorp JM, Cliver SP, Lukes A, Iams JD, Coultrip L, Eriksen N, Holbrook RH, Elliott J, Ingardia C, Pietrantoni M. Fetal Fibronectin as a predictor of preterm birth in patients with symptoms: A multicenter trial. Am J Obstet Gynecol 1997; 177: 13–18. 15. Honest H, Bachmann LM, Gupta JK, Kleijnen J, Khan KS. Accuracy of cervicovaginal fetalfibronectin test in predicting risk of spontaneous preterm birth: systematic review. BMJ 2002; 325: 301. 16. Gyetvai K, Hannah ME, Hodnett ED, Ohlsson A. Tocolytics for preterm labor: a systematic review. Obstet Gynecol 1999; 94: 869–877. 17. Joffe GM, Jacques D, Bemis-Heys R, Burton R, Skram B, Shelburne P. Impact of the fetal fibronectin assay on admissions for preterm labor. Am J Obstet Gynecol 1999; 180: 581–586. 18. Kuhrt K, Smout E, Hezelgrave N, Seed PT, Carter J, Shennan AH. Development and validation of a tool incorporating cervical length and quantitative fetal fibronectin to predict spontaneous preterm birth in asymptomatic high-risk women. Ultrasound Obstet Gynecol 2016; 47: 104–109.
SUPPORTING INFORMATION ON THE INTERNET The following supporting information may be found in the online version of this article: Appendix S1 Parameters and formulae used to calculate the probability of spontaneous preterm delivery.
Download the App QUiPP - A tool to predict spontaneous preterm birth, incorporating fetal fibronectin and cervical length, in symptomatic women and high-risk asymptomatic women. For more information go to: www.quipp.org
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Development and validation of a tool incorporating quantitative fetal fibronectin to predict spontaneous preterm birth in symptomatic women K. KUHRT, N. HEZELGRAVE, C. FOSTER, P. T. SEED and A. H. SHENNAN Woman’s Health Academic Centre, King’s College London, London, UK
K E Y W O R D S: predictive model; premature labor; quantitative fetal fibronectin; symptomatic women
ABSTRACT Objective To develop a reliable and validated tool for prediction of spontaneous preterm birth (sPTB) in symptomatic women that incorporates quantitative measurements of fetal fibronectin (qfFN) and other relevant risk factors. Methods Data were analyzed that had been collected prospectively from 382 women who presented at an emergency assessment unit between 22 + 0 and 35 + 6 weeks’ gestation with symptoms of preterm labor. Clinicians were blinded to qfFN although they were aware of qualitative fFN results. Parametric survival models for sPTB, with time-updated covariates, were developed for combinations of predictors and the best was selected using the Akaike and Bayesian information criteria. The model was developed on the first 190 consecutive women and validated on the subsequent 192. The estimated probability of delivery before 30, 34 or 37 weeks’ gestation and within 2 or 4 weeks of testing was calculated for each patient and was compared to actual event rates. Predictive statistics were calculated to compare training and validation sets. Results The final model that was selected used qfFN and previous sPTB/preterm prelabor rupture of membranes (PPROM) as predictors. Predictive statistics were similar for training and validation sets and there was good agreement between expected and observed sPTB for all outcomes. Areas under the receiver–operating characteristics curves ranged from 0.77 to 0.88, indicating accurate prediction across all five delivery outcomes. Conclusions sPTB in symptomatic women can be predicted accurately using a model combining qfFN and previous sPTB/PPROM. Clinicians can use this model, which has been incorporated into an App (QUiPP),
to determine accurately a woman’s risk of sPTB and potentially tailor management decisions appropriately. Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
INTRODUCTION Every year 15 million babies are born preterm (< 37 weeks’ gestation)1 worldwide, and 1.1 million die of associated complications2 . Survivors face significant risk of morbidity. Preterm labor is now one of the most costly indications for obstetric admission3 , but most who present with symptoms of preterm labor do not go on to deliver early4 . Due to the limited universal predictive tools available to identify those at highest risk, these women may receive unnecessary interventions, including tocolysis, steroids and admission to hospital, which are expensive and may be associated with detrimental side effects. Accurate prediction of preterm birth in symptomatic women would allow timely and appropriate intervention, potentially reducing neonatal morbidity and mortality. Fetal fibronectin (fFN) has proven a useful predictor of preterm birth in both symptomatic and high-risk asymptomatic women5,6 . fFN is an adhesive glycoprotein found in the amniotic fluid, placental tissue and extracellular component of the decidua basalis adjacent to the placental intervillous space. It is released after mechanical or inflammatory-mediated damage to the placenta or membranes before birth7 . While traditionally a binary test is used for measurement of fFN, which gives a positive or negative result based on a threshold concentration of 50 ng/mL, a novel bedside quantitative fetal fibronectin (qfFN) test has been developed that adds value over the former qualitative measurement. In our previous prospective study of 300 women at 22–35 weeks’ gestation5 , the positive predictive value
Correspondence to: Prof. A. Shennan, Division of Women’s Health, 10th Floor North Wing, St Thomas’ Hospital, London SE1 7EH, UK (e-mail: [email protected]) Accepted: 5 May 2015
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
ORIGINAL PAPER
Predicting preterm birth using quantitative fFN (PPV) for spontaneous preterm birth (sPTB; < 34 weeks’ gestation) improved with increasing fFN thresholds, from 19% to 32%, 61% and 75% at respective fFN thresholds of 10, 50, 200 and 500 ng/mL. In this way, we demonstrated that qfFN provides additional thresholds to define risk compared to the conventional qualitative test. This facilitates improved discrimination of risk and subsequent targeted management of sPTB in symptomatic women. We aimed to produce a reliable and validated tool for determining the risk of sPTB in symptomatic women attending an emergency assessment unit, using a combination of risk factors and qfFN measurement; not limited by fixed thresholds, the tool should be capable of estimating the individual probability of preterm delivery at any given gestational age or within any given number of weeks after testing. The algorithm has been incorporated into an App (QUiPP) for widespread use.
METHODS This was a prospective observational secondary analysis of a population of women enrolled in the ongoing Evaluation of Fetal Fibronectin with a Quantitative Instrument for the Prediction of Preterm Birth (EQUIPP) study involving 382 consecutive women with a singleton pregnancy who presented to an emergency assessment unit between 22 + 0 and 35 + 6 weeks’ gestation with symptoms of threatened preterm labor, as assessed by the attending clinician. Women with a blood-stained swab or sexual intercourse within the last 24 h were excluded from the study due to known interference with fFN measurement. Women with multiple pregnancy or those with insufficient or absent qfFN sample, incomplete outcome data or offspring from the current pregnancy with a major congenital abnormality were also excluded. Participants were recruited from five hospitals in the UK between October 2010 and February 2014. The study was approved by South East London Research Ethics Committee, and the local research ethics committees of all participating centers. Written informed consent was obtained from all participants. The fFN sample was obtained by an obstetrician during a sterile speculum examination in which a polyester swab was inserted into the posterior fornix of the vagina to collect cervicovaginal secretions. One aliquot (200 μL) of the sample was analyzed with the conventional fFN TLiIQ analyzer (Hologic, Marlborough, MA, USA) and a second aliquot (200 μL) of the same sample using the new quantitative Rapid fFN 10Q analyzer (Hologic). Clinicians were made aware of the qualitative TLiIQ result (positive/negative), but both patient and clinician remained blinded to 10Q results until after delivery. Participants’ demographic information, risk factors for preterm birth and obstetric and gynecological history were entered into a secure online database (www.medscinet.net/PTBstudies). Women with a positive TLiIQ fFN result were managed as per unit protocols that included administration of steroids, tocolysis and bed rest.
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
211
Any model performs optimally on the dataset that has been used to generate that model. In order to provide a true assessment of the efficacy of the model, the study population was split into a training set, consisting of 190 women with 208 visits, that was used to develop the model, and a validation set, which included 192 women with 221 visits. A predictive model was run using the validation set to assess the performance of the algorithm.
Statistical analysis Data sets Statistical analysis was performed with Stata software (version 11.2; StataCorp LP, College Station, TX, USA). The data were randomly split 1:1 into training and validation sets, each woman providing a fFN sample on one or more occasion. All test results from all participants were included in the analysis. Model generation Survival analysis with time-updated covariates for preterm birth was used to identify the principal predictors. Women were considered at risk of an event from the time of the visit to the earliest of next test, delivery or 37 weeks’ gestation. Deliveries after 37 weeks or iatrogenic preterm deliveries were regarded as censored (n = 11). fFN, gestational age at test, previous preterm delivery/PPROM and details of symptoms, i.e. abdominal pain or threatened preterm labor (more than one contraction every 10 min for at least 30 min), were considered as possible predictors. We performed an initial analysis in which other predictive variables, including body mass index, ethnicity, smoking, previous cervical surgery and previous late miscarriage, were excluded as not significant. Six parametric survival models were compared for each combination of predictors: exponential, gamma, Gompertz, log-logistic, log-normal and Weibull. The models described the probability of early delivery throughout pregnancy, as it changed according to gestational age between test and delivery. The best survival function was determined by having the lowest values of the Akaike and Bayesian information criteria (AIC and BIC)8,9 . Predictive variables were then investigated for non-linearity using fractional polynomials10 and a series of qfFN cutpoints at standard values: 10, 20, 50, 100, 200 and 500 ng/mL. Model validation To confirm test performance, the Hosmer–Lemeshow test was used to calculate actual and expected event rates by decile of risk for delivery before 30, 34 or 37 weeks’ gestation and within 2 or 4 weeks of testing11 . Receiver–operating characteristics (ROC) curves were drawn and areas under the curve (AUC) calculated. Predictive statistics of the algorithm were calculated, using a probability above 10% as indicating a positive
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Kuhrt et al.
212 Women presenting at emergency assessment unit between 22 + 0 and 35 + 6 weeks with symptoms of threatened preterm labor (n = 504 women, 592 visits)
Excluded (n = 122 women, 163 visits): Multiple pregnancy (n = 25 women, 53 visits) Congenital abnormality (n = 18 women, 22 visits) Sexual intercourse < 24 h (n = 8 women, 9 visits) Blood-stained swab (n = 12 women, 14 visits) Insufficient or absent qfFN measurement (n = 24 women, 28 visits) Incomplete outcome data (n = 35 women, 37 visits)
Final study population after exclusion (n = 382 women, 429 visits)
Training set (n = 190 women, 208 visits)
Validation set (n = 192 women, 221 visits)
Figure 1 Flow diagram of women with threatened preterm labor illustrating the number of women involved in the study after exclusions were made according to the defined criteria. qfFN, quantitative fetal fibronectin.
test in order to compare training and validation sets. This value was chosen as being practically useful for separating low-risk women (probability of delivery around 5%) from high-risk women (probability around 20%) for the purpose of clinical management.
RESULTS After exclusions, as illustrated in Figure 1, the final population of symptomatic women studied consisted of 382 women, and 429 fFN measurements were recorded. This final group was divided into a training set, consisting of 190 women with 208 visits, and a validation set of 192 women and 221 visits. Of the study population, 73% had only a single visit to the unit, at a mean gestational age of 29 weeks; those who had two or more visits typically visited earlier in pregnancy, at a mean gestational age of 25 weeks. The demographic characteristics of the validation and trainings sets were comparable (Table 1) as was the proportion of women in each fFN concentration category (Table 2). In the total study population, 29 (8%) women received tocolysis as per management protocols, two (0.5%) received progesterone and four (1%) received cervical cerclage. The rate of sPTB was 13%, 6%, 3%, 3% and 7% for delivery at < 37, < 34 and < 30 weeks’ gestation and within 2 and 4 weeks of testing, respectively, in the training set and 13%, 7%, 5%, 5% and 10%, respectively, in the validation set. These values are comparable to rates reported in the literature1,2,5 .
Model generation (using the training set only) Of the predictors considered, only qfFN and previous sPTB/PPROM were significant in the stepwise model at P < 0.05 (Table 3). The best parametric survival model
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
Table 1 Demographic characteristics of women with singleton pregnancy presenting with threatened preterm labor in training and validation sets Demographic characteristic Age (years) Ethnicity White Black Asian Other BMI* < 20 kg/m2 20–24.9 kg/m2 25–29.9 kg/m2 ≥ 30 kg/m2 Smoker*† Never Current Stopped Before pregnancy During pregnancy Risk factors Previous sPTB Previous PPROM Previous late miscarriage (16 + 0 to 23 + 6 weeks) Previous cervical surgery
Training set (n = 190)
Validation set (n = 192)
All (n = 382)
30
30
30
92 (48) 68 (36) 11 (6) 19 (10)
92 (48) 65 (34) 20 (10) 15 (8)
184 (48) 133 (35) 31 (8) 34 (9)
27 (14) 81 (43) 41 (22) 41 (22)
32 (17) 73 (38) 48 (25) 37 (19)
59 (16) 154 (41) 89 (23) 78 (20)
138 (73) 23 (12)
148 (78) 23 (12)
286 (76) 46 (12)
19 (10) 9 (5)
15 (8) 4 (2)
34 (9) 13 (3)
36 (19) 13 (7) 14 (7)
28 (15) 9 (5) 14 (7)
64 (17) 22 (6) 28 (7)
14 (7)
12 (6)
26 (7)
Data are given as mean or n (%). There was no significant difference in cervical length between training and validation sets. *Two values missing from validation set. †One value missing from training set. BMI, body mass index; PPROM, preterm prelabor rupture of membranes; sPTB, spontaneous preterm birth.
(determined by lowest AIC and BIC) was a log-normal survival function with terms for qfFN and previous sPTB/PPROM. The best fractional polynomial model was linear for fFN concentration, outperforming those using
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Predicting preterm birth using quantitative fFN
213
Table 2 Fetal fibronectin concentration (FN) for training and validation sets comprising women with singleton pregnancy presenting with threatened preterm labor
Fetal FN
Training set (n = 190)
Validation set (n = 192)
All (n = 382)
114 (60) 20 (11) 21 (11) 10 (5) 10 (5) 15 (8)
113 (59) 19 (9) 15 (8) 9 (5) 18 (9) 18 (9)
227 (59) 39 (10) 36 (9) 19 (5) 28 (7) 33 (9)
< 10 ng/mL 10–19.9 ng/mL 20–49.9 ng/mL 50–99.9 ng/mL 100–199.9 ng/mL ≥ 200 ng/mL
Data are given as n (%). There was no significant difference in fetal fibronectin concentration between training and validation sets. Table 3 Variables tested for inclusion in the model for prediction of spontaneous preterm birth Variable qfFN Previous sPTB/PPROM Gestational age at test Abdominal pain Threatened preterm labor*
Parameter estimate
P
−0.0015 (−0.0024 to −0.0006) −0.18 (−0.35 to −0.013) 0.00038 (−0.0020 to 0.0028) −0.25 (−0.70 to 0.19) −0.16 (−0.64 to 0.31)
0.0002 0.035 0.753 0.264 0.496
Values in parentheses are 95% CIs. *More than one contraction every 10 min for at least 30 min. PPROM, previous preterm prelabor rupture of membranes; qfFN, quantitative fetal fibronectin; sPTB, spontaneous preterm birth.
cutpoints. These findings were confirmed by improvement in AIC and BIC. Therefore the final model included linear fFN and previous sPTB/PPROM (Appendix S1).
Model validation Table 4 summarizes predictive statistics calculated using an estimated probability of delivery > 10% as indicating a positive test, to allow comparison of training and validation sets for prediction of delivery at five clinically important points in time: < 37, < 34 and < 30 weeks’ gestation and within 2 or 4 weeks of qfFN testing. Predictive statistics, including sensitivity, specificity, positive (LR+) and negative (LR–) likelihood ratios were similar for training and validation sets across all five time points considered. The LR+ and LR– values were all much greater than or less than 1, respectively, which indicates that positive test results were strongly associated with the occurrence of sPTB and negative test results with its absence12 . On comparison of observed and expected sPTB rates (Table 5), test performance was generally good. No significant differences were found for women with a probability of early delivery > 10%. For women with a probability ≤ 10%, the observed rate was always < 10%, as required; but it was significantly higher than expected for delivery < 30 weeks and for delivery within 4 weeks.
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
Significance disappears after allowing for multiple testing. However, extremely low probabilities may need to be interpreted with caution. The ROC curves in Figure 2 represent overall test performance regardless of fFN threshold. The AUC values (Table 4), ranging from 0.77–0.88 in the validation set, indicate that the model provides useful information for prediction of sPTB across all five delivery time points investigated.
DISCUSSION We have created an accurate prediction model incorporating qfFN and previous sPTB/PPROM that can be used for the prediction of sPTB in symptomatic women. The model remains accurate when tested on a validation set. Its reliable performance is demonstrated further by comparing expected and observed sPTB rates which gave P-values > 0.05 consistently, indicating that any differences are not significant and are therefore unlikely to be clinically important. This is supportive of previous research that showed that quantification of fFN provides additional information on the risk of sPTB in symptomatic5 and asymptomatic high-risk13 women when compared with the original qualitative fFN test. Although overall sensitivity and specificity are similar to those in the literature, the advantage of the algorithm is that the individual risk for each woman can be determined instead of calculation of a summary value. A previous study including 725 singleton pregnancies demonstrated similar negative predictive values (NPVs) but much lower PPVs; Peaceman et al.14 reported a PPV of 16% for prediction of delivery < 34 weeks’ gestation compared to 33.3% in our study. This suggests that, using our model, clinicians can be more confident in the validity of a positive test result, and can accurately target interventions while avoiding treatment in women who do not need it. The ROC curves (Figure 2) illustrate overall prediction of sPTB regardless of qfFN threshold, which is highly relevant given the use of qfFN in this model, and will provide optimal prediction throughout the range of values on which a clinician can act. AUCs ranged from 0.77 to 0.88, showing an improvement on those reported in previously published literature; Honest et al. performed a systematic review of studies using qfFN at a threshold of 50 ng/mL. Forty studies involving 26 876 women were included and summary ROC areas ranged from 0.71 to 0.77 for prediction of delivery at < 37 and < 34 weeks’ gestation, respectively15 . A major strength of this study is that data were derived from a prospective dataset in which clinicians and patients were blinded to the qfFN result. Furthermore, the parametric method employed to generate the model, combined with appropriate censoring, makes full use of the available data without introducing biases. However, a potential limitation is the small number of women with sPTB at some gestational time points; a larger study could provide the opportunity to further validate our findings.
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Kuhrt et al.
214
Table 4 Predictive statistics for delivery at < 30, < 34 and < 37 weeks’ gestation and delivery within 2 or 4 weeks from fetal fibronectin test in training and validation datasets GA at delivery Parameter Sensitivity (%) Training Validation Specificity (%) Training Validation AUC Training Validation LR+ Training Validation LR– Training Validation PPV (%) Training Validation NPV (%) Training Validation
Delivery within:
< 30 weeks
< 34 weeks
< 37 weeks
2 weeks
4 weeks
100 (47.8–100) 50.0 (18.7–81.3)
66.7 (34.9–90.1) 78.6 (49.2–95.3)
66.7 (44.7–84.4) 72.0 (50.6–87.9)
44.4 (13.7–78.8) 50.0 (18.7–81.3)
68.4 (43.4–87.4) 47.4 (24.4–71.1)
97.8 (94.6–99.4) 96.2 (92.2–98.4)
90.4 (85.1–94.3) 87.6 (81.9–92.1)
78.9 (71.9–84.9) 77.2 (70.1–83.4)
93.4 (88.7–96.5) 94.0 (89.4–96.9)
90.6 (85.3–94.6) 90.2 (84.7–94.2)
1.0 (1.0–1.0) 0.88 (0.74–1.0)
0.76 (0.63–0.89) 0.83 (0.67–0.98)
0.76 (0.63–0.89) 0.77 (0.65–0.90)
0.70 (0.43–0.98) 0.77 (0.55–0.99)
0.80 (0.65–0.95) 0.78 (0.66–0.91)
46.3 (17.5–122.0) 13.0 (5.0–33.8)
7.0 (3.8–12.8) 6.4 (3.9–10.3)
3.2 (2.1–4.8) 3.2 (2.2–4.6)
6.7 (2.7–16.7) 8.3 (3.6–19.2)
7.3 (4.2–12.8) 4.8 (2.5–9.3)
0.09 (0.01–1.21)* 0.5 (0.3–1.0)
0.4 (0.2–0.8) 0.2 (0.1–0.7)
0.4 (0.2–0.8) 0.4 (0.2–0.7)
0.6 (0.3–1.1) 0.5 (0.3–1.0)
0.4 (0.2–0.7) 0.6 (0.4–0.9)
55.5 (21.2–86.3) 41.7 (15.2–72.3)
32.0 (14.9–53.5) 33.3 (18.0–51.8)
31.4 (19.1–45.9) 32.1 (20.3–46.0)
25.0 (7.3–52.4) 31.3 (11.0–58.7)
44.8 (26.4–64.3) 34.6 (17.2–55.7)
100 (98.0–100) 97.2 (93.6–99.1)
97.6 (93.9–99.3) 98.1 (94.6–99.6)
94.2 (89.0–97.5) 94.9 (89.7–97.9)
97.1 (93.4–99.1) 97.2 (93.5–99.1)
96.3 (92.1–98.6) 94.0 (89.2–97.1)
For each gestational time point, an individual probability of early delivery > 10% was treated as a positive test result. Values in parentheses are 95% CIs. *Calculated using substitution formula: 0.5 added to allow for zero values. AUC, area under receiver–operating characteristics curve; GA, gestational age; LR–, negative likelihood ratio; LR+, positive likelihood ratio; NPV, negative predictive value; PPV, positive predictive value. Table 5 Expected and observed rates of spontaneous preterm birth at < 30, < 34 and < 37 weeks’ gestation and within 2 or 4 weeks from fetal fibronectin testing in the validation set (n = 192)
Time point of delivery < 30 weeks P ≤ 10 P > 10 < 34 weeks P ≤ 10 P > 10 < 37 weeks P ≤ 10 P > 10 Within 2 weeks P ≤ 10 P > 10 Within 4 weeks P ≤ 10 P > 10
Event rate (n/N)
Expected rate* (%)
Observed rate (95% CI) (%)
P†
5/180 5/12
0.79 54
2.8 (0.9–6.4) 42 (15–72)
0.02 0.40
3/159 11/33
2.1 40
1.9 (0.4–5.4) 33 (18–52)
1.00 0.49
7/136 18/56
4.5 39
5.2 (2.1–10.3) 32 (20.3–46.1)
0.68 0.34
5/176 5/16
1.3 27
2.8 (0.9–6.5) 31 (11.0–58.7)
0.08 0.78
10/166 9/26
2.5 36
6 (2.9–10.8) 35 (17.2–55.7)
0.01 1.00
For each gestational time point, an individual probability of early delivery ≤ 10% (P ≤ 10) or > 10% (P > 10) was treated as a negative or positive test result, respectively. *Calculated using algorithm. †Comparison of observed and expected rates.
The low rate of tocolysis in the study population likely reflects the practice applied at the units from which patients were recruited, as it is generally used only for in-utero transfer. The small numbers involved are unlikely to impact significantly on sPTB prediction. However, to
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
clarify, 8% of the total study population received tocolysis as per management protocols, which has been shown to prolong pregnancy for 48 h to 7 days16 and 0.5% and 1% received progesterone or cerclage, respectively. Evaluation of the predictive value of fFN independent of decisions regarding these interventions is difficult in current clinical practice, and our prediction model takes this management into account. Reactive therapies such as magnesium sulphate and steroids can still be targeted, but further investigation is required to ascertain how the model predicts without tocolysis, progesterone and cerclage, i.e. clinical evaluation of the model in practice is needed. It is not common practice in the UK to use cervical length as a predictor in the acute setting for symptomatic threatened preterm birth, and we do not have systematic data to include in the algorithm. Future work should incorporate both cervical length and qfFN to obtain a potentially improved model. In the meantime, it remains reasonable to use digital examination and cervical change to ascertain risk of delivery, to confirm true labor and to rule out the rare false-negative fFN test result when clinical suspicion dictates. This subjective assessment has not undergone formal evaluation as a predictive tool and is likely to be superseded in the future by cervical length scanning, particularly in early labor. In practice, the high NPV demonstrated for all gestational time points investigated will enable clinicians confidently to discharge women deemed to be at low risk by the model, without administering costly interventions
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Predicting preterm birth using quantitative fFN
215
0.75
0.75
0.75
0.50
0.25
0
Sensitivity
(c) 1.00
Sensitivity
(b) 1.00
Sensitivity
(a) 1.00
0.50
0.25
0.25
0.25
0.50
0.75
0
1.00
0.25
(e) 1.00
0.75
0.75 Sensitivity
(d) 1.00
Sensitivity
0.50
0.75
1.00
0.50
0.50
0.75
1.00
0.50
0
0.25
1 – Specificity
0.75
1.00
1 – Specificity
associated with potential side effects. Furthermore, the strong overall prediction illustrated by the ROC curves will enable management decisions regarding in utero transfer, tocolytic therapy, antenatal corticosteroids and magnesium sulphate to be tailored appropriately to suit each woman, based on her risk. There is already evidence to show that qualitative fFN sampling in symptomatic women reduces the number of admissions, length of stay and administration of tocolysis. One study demonstrated a 40% reduction in hospital admissions for preterm labor and cost savings of approximately US$486 000, with no negative impact on neonatal outcome17 . With the added value from qfFN testing, admissions and unnecessary interventions could be reduced and costs cut even further. More work is needed to ascertain whether interventions improve outcome when targeted to women identified by the algorithm, but some interventions may be withheld in women deemed to be low risk by the algorithm. The addition of cervical length measurement may further improve prediction and this should be investigated in future studies; we were unable to incorporate it into the current algorithm as we do not have those data. However, another algorithm has been developed for asymptomatic women incorporating qfFN and cervical length18 . Together, these data form the basis of a model for prediction of sPTB in asymptomatic high-risk and symptomatic women for whom the risk of delivery at five clinically important points in time can be calculated accurately based on the qfFN result and previous history of sPTB/PPROM. The algorithm is available through an App (QUiPP) to make it easily accessible to clinicians for use at the bedside and in clinic.
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
0.50
0.25
0.50
0.75
1.00
1 – Specificity
0.25
0.25
0.25
0
1 – Specificity
1 – Specificity
0
0.50
Figure 2 Receiver operating–characteristics (ROC) curves showing overall prediction for delivery < 30 weeks (a), < 34 weeks (b) and < 37 weeks (c), and within 2 weeks (d) or 4 weeks (e) of testing in the validation set, for a model including linear fetal fibronectin and previous spontaneous preterm birth or preterm prelabor rupture of membranes. Areas under ROC curve: (a) 0.8816, (b) 0.8262, (c) 0.7728, (d) 0.7676 and (e) 0.7837.
ACKNOWLEDGMENTS We thank the women who participated in this study, Judy Filmer and the team who assisted with patient recruitment and sample processing from the Preterm Surveillance Clinic at the Division of Women’s Health, King’s College London. This research was supported by Tommy’s Baby Charity (charity number 1060508)
DISCLOSURES Minority financial and equipment assistance was provided by Hologic USA (Marlborough, MA, USA). A.H.S. has received financial assistance from Hologic USA for providing educational talks on preterm birth.
REFERENCES 1. Blencowe H, Cousens S, Oestergaard MZ, Chou D, Moller AB, Narwal R, Adler A, Vera Garcia C, Rohde S, Say L, Lawn JE. National, regional, and worldwide estimates of preterm birth rates in the year 2010 with time trends since 1990 for selected countries: a systematic analysis and implications. Lancet 2012; 379: 2162–2172. 2. March of Dimes, PMNCH, Save the Children, WHO. Born Too Soon: The global action report on preterm birth. Howson CP, Kinney MV, Lawn JE (eds). World Health Organization: Geneva, 2012. http://www.who.int [accessed 29 February 2014]. 3. Russell RB, Green NS, Steiner CA, Meikle S, Howse JL, Poschman K, Dias T, Potetz L, Davidoff MJ, Damus K, Petrini JR. Cost of hospitalization for preterm and low birth weight infants in the United States. Pediatrics 2007; 120: e1–9. 4. Guinn DA, Goepfert AR, Owen J, Brumfield C, Hauth JC. Management options in women with preterm uterine contractions: a randomized clinical trial. Am J Obstet Gynecol 1997; 177: 814–818. 5. Abbott DS, Radford SK, Seed PT, Tribe RM, Shennan AH. Evaluation of quantitative fetal fibronectin test for spontaneous preterm birth in symptomatic women. Am J Obstet Gynecol 2012; 208: 122.e1–122.e6. 6. Bolt LA, Chandiramani M, de Greeff A, Seed PT, Kurtzman J, Shennan AH. The value of combined cervical length measurement and fetal fibronectin testing to predict spontaneous preterm birth in asymptomatic high risk women. J Matern Fetal Neonatal Med 2011; 24: 928–932.
Ultrasound Obstet Gynecol 2016; 47: 210–216.
Kuhrt et al.
216 7. Genc MR, Ford CE. The clinical use of inflammatory markers during pregnancy. Curr Opin Obstet Gynecol 2010; 22: 116–121. 8. Akaike H. Information theory and the extension of the maximum likelihood principle. In Second International Symposium on Information Theory. Petrov VN, Csaki F (eds), Akailseoniai-Kiudo: Budapest, 1973; 267–281. 9. Schwarz G. Estimating the dimension of a model. Ann Stat 1978; 6: 461–464. 10. Sauerbrei W & Royston P. Building multivariable prognostic and diagnostic models: Transformation of the predictors by using fractional polynomials. J R Stat Soc Ser A Stat Soc 1999; 162: 71–94. 11. Hosmer DW Jr, Lemeshow S, Klar J. Goodness-of-fit testing for the logistic regression model when the estimated probabilities are small. Biom J 1988; 30: 911–924. 12. Deeks JJ, Altman DG. Diagnostic tests 4: likelihood ratios. BMJ 2004; 329: 168–169. 13. Kurtzman J, Chandaramani M, Briley A, Poston L, Das A, Shennan AH. Quantitative fetal fibronectin in asymptomatic high risk patients and the spectrum of risk for recurrent preterm delivery. Am J Obstet Gynecol 2009; 200: 263.e1–263.e6.
14. Peaceman AM, Andrews WW, Thorp JM, Cliver SP, Lukes A, Iams JD, Coultrip L, Eriksen N, Holbrook RH, Elliott J, Ingardia C, Pietrantoni M. Fetal Fibronectin as a predictor of preterm birth in patients with symptoms: A multicenter trial. Am J Obstet Gynecol 1997; 177: 13–18. 15. Honest H, Bachmann LM, Gupta JK, Kleijnen J, Khan KS. Accuracy of cervicovaginal fetalfibronectin test in predicting risk of spontaneous preterm birth: systematic review. BMJ 2002; 325: 301. 16. Gyetvai K, Hannah ME, Hodnett ED, Ohlsson A. Tocolytics for preterm labor: a systematic review. Obstet Gynecol 1999; 94: 869–877. 17. Joffe GM, Jacques D, Bemis-Heys R, Burton R, Skram B, Shelburne P. Impact of the fetal fibronectin assay on admissions for preterm labor. Am J Obstet Gynecol 1999; 180: 581–586. 18. Kuhrt K, Smout E, Hezelgrave N, Seed PT, Carter J, Shennan AH. Development and validation of a tool incorporating cervical length and quantitative fetal fibronectin to predict spontaneous preterm birth in asymptomatic high-risk women. Ultrasound Obstet Gynecol 2016; 47: 104–109.
SUPPORTING INFORMATION ON THE INTERNET The following supporting information may be found in the online version of this article: Appendix S1 Parameters and formulae used to calculate the probability of spontaneous preterm delivery.
Download the App QUiPP - A tool to predict spontaneous preterm birth, incorporating fetal fibronectin and cervical length, in symptomatic women and high-risk asymptomatic women. For more information go to: www.quipp.org
Copyright © 2015 ISUOG. Published by John Wiley & Sons Ltd.
Ultrasound Obstet Gynecol 2016; 47: 210–216.
