Predictive performance of cervical elastography for preterm birth: A head-to-head network meta-analysis

Abstract

BACKGROUND

Early detection of cervical remodeling is vital for predicting preterm birth (PTB). Conventional morphologic indices, such as cervical length (CL) and anterior cervical angle (ACA), demonstrate limited diagnostic accuracy. Elastic imaging of the cervix (E-cervix) elastography offers functional biomechanical insights by quantifying cervical strain and stiffness. However, a systematic comparison of its diverse parameters alongside traditional indices is currently lacking. Identifying the most effective ultrasound-based markers is essential for improving mid-trimester risk stratification and facilitating timely clinical interventions to improve perinatal outcomes.

AIM

To evaluate and compare the diagnostic performance of E-cervix parameters, CL, and ACA for PTB.

METHODS

PubMed, Web of Science, EMBASE, Cochrane Library, China National Knowledge Infrastructure, and Wanfang Data were searched up to December 1, 2025. Studies reporting diagnostic accuracy of E-cervix parameters, CL, or ACA for PTB were included. A Bayesian network meta-analysis estimated pooled sensitivity, specificity, and diagnostic odds ratios (DOR). The comparative performance was ranked using the advantage index. Heterogeneity was assessed via Cochran’s Q and I² statistics, with meta-regression exploring potential modifiers.

RESULTS

Fifteen studies involving 1965 pregnancies (570 PTB cases) were included. The multi-parametric combined model demonstrated the highest diagnostic performance, with an area under the curve of 0.91, sensitivity of 0.83, specificity of 0.83, and a DOR of 26.23. Among individual parameters, ACA (DOR: 13.65) and the E-cervix hardness ratio (DOR: 7.71) significantly outperformed conventional CL (DOR: 3.58). Subgroup analysis revealed that the combined model was particularly effective in singleton pregnancies (sensitivity: 0.837, specificity: 0.882). No significant publication bias was detected except for elasticity contrast index.

CONCLUSION

E-cervix elastography provides superior diagnostic value compared to conventional CL measurement. The integration of biomechanical and morphological data via multi-parametric models offers the most precise risk stratification for PTB.

Key Words: Preterm birth; Cervical length; Cervical elastography; Elastic imaging of the cervix; Hardness ratio; Network meta-analysis

Core Tip: Cervical length and anterior cervical angle are widely used to assess preterm birth risk; however, they overlook early biomechanical remodeling. By synthesizing 15 studies, this head-to-head Bayesian network meta-analysis demonstrates that cervical elastography, particularly the hardness ratio, enhances discrimination, and that combined morphologic-biomechanical models achieve the best overall accuracy. These results support the integration of cervical elastography into risk stratification and highlight the need for standardized protocols and thresholds.

INTRODUCTION

Preterm birth (PTB) remains a global leading cause of perinatal morbidity[1]. Early identification of high-risk women allows for timely interventions, such as progesterone or cerclage, which significantly improve outcomes[2]. Consequently, accurate mid-trimester risk stratification is a critical challenge in perinatal medicine.

Transvaginal ultrasound measurement of cervical length (CL) is the most widely used tool for predicting PTB[3]. The anterior cervical angle (ACA) is increasingly employed as a complementary parameter. However, CL shows only moderate sensitivity (30%-80%) with a high false-negative rate (20%-70%)[3-5]. ACA lacks standardized measurement techniques and diagnostic thresholds[6]. Both are mostly morphological indices and may miss early cervical remodeling and softening.

Elastic imaging of the cervix (E-cervix) has emerged as a functional ultrasound technique to quantitatively assess tissue strain and biomechanical properties. By detecting cervical softening before structural shortening occurs, E-cervix may enable more precise risk assessment[7]. Single-center studies have suggested that E-cervix outperforms CL or ACA, although findings were often limited by small samples and heterogeneous parameter definitions. To date, no diagnostic network meta-analysis has comprehensively compared the relative performance of diverse E-cervix parameters with CL and ACA within the same populations.

Therefore, this study systematically evaluated the diagnostic accuracy of E-cervix elastography for predicting PTB and directly compared its performance with CL and ACA in head-to-head analyses. In addition, a multivariable diagnostic test accuracy network meta-analysis was performed to compare different E-cervix parameters with CL, providing robust evidence to support the integration of E-cervix into PTB risk assessment and clinical decision-making.

MATERIALS AND METHODS

Search strategy

A standardized literature search was performed from the earliest reports on E-cervix up to December 1, 2025 in PubMed, Web of Science Core Collection, EMBASE, the Cochrane Library, China National Knowledge Infrastructure, and Wanfang Data, limited to studies published in English or Chinese. The search combined Medical Subject Headings terms and free-text keywords related to PTB (e.g., ‘preterm birth’, ‘premature birth’) and cervical elastography (e.g., ‘E-cervix’, ‘cervical elastography’, ‘cervical strain’), using Boolean operators; the full strategy is provided in Supplementary Table 1. In addition, we screened the reference lists of relevant articles and reviews. For studies with missing data, corresponding authors were contacted via email; although two authors responded, only one provided the relevant information[8,9].

Inclusion and exclusion criteria

We included prospective or retrospective studies that evaluated the diagnostic accuracy of E-cervix elastography for predicting PTB in pregnant women (any gestational age at testing). PTB was confirmed by gestational age at delivery according to the definition of each study (details summarized in Table 1). Studies were required to report, or allow reconstruction of, a 2 × 2 table for at least one E-cervix parameter and for the corresponding CL and/or ACA when available. We excluded studies without sufficient data to derive sensitivity and specificity, those assessing only CL or only ACA without E-cervix, conference abstracts, reviews, case reports, animal or in-vitro studies, unpublished data, and duplicate or overlapping cohorts (retaining the most complete or recent report).

Table 1 Characteristics of the included studies (n = 15).

Ref.	Country	Study design	Pregnancy status	Gestational weeks	Mean age (in years)	Sample size (PTB/TD)	Detection method	Ultrasonic diagnostic device
Zhang et al[14], 2021	China	Retrospective	Singleton pregnancy	16~28	22~41 (30.01 ± 4.03)	171 (24/147)	CL, IOS, IOS/EOS ratio, HR	Samsung Medison WS80A
Xiong et al[15], 2021	China	Retrospective	Singleton pregnancy	20~24	PTB: 26 (23.75, 27.00); TD: 26 (25.00, 27.00)	85 (40/45)	CL, IOS, ECI, EOS	Samsung Medison WS80A
Wang et al[16], 2022	China	Retrospective	Singleton pregnancy	14~28	20~36 (29.96 ± 3.01)	200 (31/169)	CL, ECI, IOS, IOS/EOS ratio, HR	Samsung Medison WS80A
Cha et al[17], 2022	Korea	Prospective	Singleton pregnancy	14~24	PTB: 33.8 ± 3.8; TD: 34.2 ± 3.8	66 (8/58)	CL, ECI, EOS, IOS, HR	Samsung Medison WS80A
Nazzaro et al[18], 2022	Italy	Retrospective	Singleton pregnancy	23~33	PTB: 28.9 ± 6.1; TD: 28.4 ± 6.1	95 (42/53)	CL, HR	Samsung Medison WS80A
Li et al[19], 2023	China	Retrospective	Singleton Pregnancy	14~27 and 28~36	PTB: 34 (29, 37); TD: 31 (28, 32)	149 (31/118)	HR	Samsung HERA W10/W9
Li et al[20], 2023	China	Retrospective	Singleton pregnancy	16~28 and 22~24	PTB: 24~37 (29.7 ± 3.6); TD: 24~37 (29.6 ± 3.6)	82 (14/68)	ACA, ECI, IOS/EOS ratio, HR	Samsung (NR)
Wu et al[21], 2024	China	Retrospective	Singleton and multiple pregnancy	18~27	23~40 (30.22 ± 4.03)	225 (105/120)	ECI, EOS, IOS	Samsung HERA W10
Liu et al[22], 2024	China	Prospective	Multiple pregnancy	18~24	21~40 (30.9 ± 3.8)	97 (35/62)	ACA, ECI, IOS, EOS, IOS/EOS ratio, HR	Samsung HERA W9
He et al[23], 2024	China	Prospective	Singleton pregnancy	18~24	PTB: 30.6 ± 3.2; TD: 30.2 ± 3.3	228 (26/202)	CL, EOS, IOS	Samsung Medison WS80A
Nazzaro et al[24], 2024	Italy	Retrospective	Multiple pregnancy	23~33	PTB: 27.9 ± 5.5; TD: 29.3 ± 6.2	63 (44/19)	CL, HR	Samsung Medison WS80A
Yu et al[25], 2025	China	Prospective	Multiple pregnancy	20~24	18~40 (32.50 ± 3.27)	130 (40/90)	CL, IOS, HR	Samsung HERA XW9
He et al[26], 2025	China	Retrospective	Singleton Pregnancy	16~28	22~40 (31.12 ± 4.06)	88 (21/67)	ACA, ECI, HR	Samsung (NR)
Ma et al[27], 2025	China	Retrospective	Singleton pregnancy	11~18	PTB: 29.10 ± 2.39; TD: 29.12 ± 2.36	136 (25/111)	CL, ACA, ECI, EOS, IOS, IOS/EOS ratio, HR	Samsung (NR)
Kwon et al[28], 2025	Korea	Prospective	Singleton and multiple pregnancy	24~34	PTB:34 (21, 40); TD: 33 (26, 42)	150 (84/66)	CL, ECI, IOS/EOS ratio	Samsung Medison WS80A

PTB: Preterm birth; TD: Term delivery; CL: Cervical length; ACA: Anterior cervical angle; ECI: Elasticity contrast index; IOS: Internal os strain; EOS: External os strain; HR: Hardness ratio.

Study selection and data extraction

Two researchers (Zhang HP and Li N) independently screened titles and abstracts, followed by a full-text assessment conducted by the other two researchers (Li F and Zhou YQ). Disagreements were resolved by consensus or by consulting a third reviewer within the team. Data extraction was performed by two investigators (Zhang HP and Li N) for eight parameters: CL, ACA, and five E-cervix indices - elasticity contrast index (ECI), internal os strain (IOS), external os strain (EOS), internal-to-EOS ratio (IOS/EOS ratio), and hardness ratio (HR). Additionally, we extracted data for “combined” models, which were defined as the integration of at least two E-cervix indices, or the combination of at least one E-cervix index with CL and/or ACA.

Risk of bias assessment

Two investigators assessed methodological quality using Quality Assessment of Diagnostic Accuracy Studies-2 across patient selection, index test, reference standard, and flow/timing. Disagreements were resolved by discussion with a third reviewer. Risk of bias and applicability concerns were rated as low, high, or unclear[10].

Statistical analysis

Statistical analyses were performed using Meta-DiSc 1.4, Stata 17.0, and R 4.3.1, with P < 0.05 considered statistically significant unless otherwise specified. Primary diagnostic performance was evaluated by extracting data from 2 × 2 tables to calculate pooled sensitivity, specificity, likelihood ratios (likelihood ratios+/Likelihood ratios-), and diagnostic odds ratio (DOR) through bivariate random-effects models and summary receiver operating characteristic curves. To evaluate relative performance, a Bayesian multivariate network meta-analysis was conducted in R, accounting for within-study correlations and between-study heterogeneity, with CL serving as the reference standard and all parameters ranked by the advantage index[11,12]. Heterogeneity was assessed via Spearman correlation for threshold effects (P > 0.05 indicating no effect) and Cochran’s Q and I² statistics for non-threshold heterogeneity. Potential sources were further explored through meta-regression and subgroup analyses involving pregnancy status, gestational age, and study design[13]. Publication bias was examined using Deeks’ funnel plot asymmetry test (P < 0.10). Finally, the robustness of the primary findings was confirmed through sensitivity analyses using Stata’s midas framework, and the methodology was independently reviewed by Li F.

RESULTS

Study selection and characteristics

From 69 identified records, 46 remained after removing duplicates. Initial screening led to the retrieval of 37 full-text papers, and 22 were excluded for failing to meet eligibility criteria. Ultimately, 15 studies (2021-2025) involving 1965 pregnancies (570 PTB, 29.0%) were included (Figure 1). Twelve studies were conducted in China, two in South Korea, and one in Italy. Five studies were prospective, and 10 were retrospective. Gestational age at screening ranged from 11 weeks to 34 weeks. All studies utilized Samsung ultrasound platforms (Medison WS80A and HERA series, Samsung Medison, South Korea). Frequently assessed indices included HR (11 studies), ECI (10), IOS (9), EOS (6), and the IOS/EOS ratio (6). CL (11 studies) and ACA (4 studies) served as conventional comparators (Table 1)[14-28].

Figure 1 PRISMA flowchart outlining the literature screening process, study selection, and exclusion criteria.

Quality assessment of the included studies

The Quality Assessment of Diagnostic Accuracy Studies-2 assessment (Figure 2) revealed no high risk of bias across any domain. All studies showed low risk for reference standard and flow and timing, reflecting consistent PTB definitions and adequate follow-up. Risk of bias for patient selection and index tests was primarily low or unclear, the latter due to insufficient reporting on consecutive enrollment or blinding. Applicability concerns were rated as low across all domains.

Figure 2 Quality Assessment of Diagnostic Accuracy Studies-2 criteria for the included studies. ¹Distinct independent study from Li et al[20] in 2023 with the same first author surname ‘Li et al[19]’.

Comparative diagnostic accuracy

Network meta-analysis evaluated eight ultrasound indices: CL, ACA, ECI, IOS, EOS, IOS/EOS ratio, HR, and a combined multi-parameter model (Table 2). Among conventional measures, CL showed moderate accuracy [sensitivity = 0.63, 95% credible interval (CrI): 0.54-0.71; specificity = 0.67, 95%CrI: 0.57-0.76; DOR = 3.58, 95%CrI: 1.91-6.25]. ACA performed substantially better (sensitivity = 0.78, 95%CrI: 0.65-0.89; specificity = 0.75, 95%CrI: 0.58-0.87; DOR = 13.65, 95%CrI: 3.88-35.20).

Table 2 Combined area under the curve, sensitivity, specificity, and diagnostic odds ratio for E-cervix parameters and conventional measurement parameters.

Parameters	AUC (95%CrI)	Sensitivity (95%CrI)	Specificity (95%CrI)	DOR (95%CrI)	LR+ (95%CrI)	LR- (95%CrI)	S	RR_Se vs CL	RR_Sp vs CL
CL	0.70 (0.60-0.77)	0.63 (0.54-0.71)	0.67 (0.57-0.76)	3.58 (1.91-6.25)	1.89 (1.24-2.90)	0.56 (0.39-0.82)	0.28 (0.08-1.00)	1.00 (1.00-1.00)	1.00 (1.00-1.00)
ACA	0.86 (0.71-0.93)	0.78 (0.65-0.89)	0.75 (0.58-0.87)	13.65 (3.88-35.20)	3.15 (1.55-7.04)	0.29 (0.13-0.61)	5.70 (0.33-15.00)	1.25 (1.01-1.53)	1.13 (0.86-1.40)
IOS	0.76 (0.65-0.83)	0.56 (0.45-0.67)	0.80 (0.70-0.87)	5.47 (2.59-10.06)	2.75 (1.50-5.13)	0.55 (0.38-0.78)	0.73 (0.14-3.00)	0.90 (0.71-1.10)	1.20 (1.01-1.44)
EOS	0.72 (0.68-0.76)	0.57 (0.51-0.64)	0.78 (0.75-0.81)	4.00 (2.81-5.69)	2.27 (1.65-3.12)	0.61 (0.50-0.75)	0.58 (0.08-2.33)	1.07 (0.86-1.29)	0.98 (0.75-1.24)
IOS/EOS ratio	0.72 (0.58-0.80)	0.68 (0.57-0.78)	0.63 (0.49-0.75)	4.05 (1.64-8.14)	1.84 (1.10-3.17)	0.50 (0.30-0.89)	0.46 (0.08-2.33)	1.09 (0.89-1.31)	0.95 (0.72-1.19)
ECI	0.73 (0.63-0.80)	0.67 (0.57-0.76)	0.67 (0.56-0.77)	4.52 (2.24-8.21)	2.04 (1.30-3.29)	0.49 (0.31-0.77)	0.60 (0.09-2.33)	1.08 (0.89-1.29)	1.01 (0.81-1.23)
HR	0.80 (0.72-0.85)	0.76 (0.68-0.82)	0.70 (0.60-0.78)	7.71 (4.06-13.06)	2.51 (1.69-3.75)	0.35 (0.23-0.54)	2.18 (0.20-9.00)	1.21 (1.04-1.41)	1.05 (0.87-1.28)
Combined	0.91 (0.83-0.95)	0.83 (0.74-0.89)	0.83 (0.73-0.90)	26.23 (10.47-54.64)	4.76 (2.70-8.74)	0.21 (0.12-0.36)	12.90 (3.67-15.00)	1.32 (1.13-1.55)	1.25 (1.04-1.48)

Heterogeneity was quantified using Cochran’s Q test and the I² statistic; a random-effects model was employed if I² > 50% or P < 0.10 for Cochran’s Q test, indicating significant heterogeneity; otherwise, a fixed-effects model was used. AUC: Area under the curve; CrI: Credible interval; DOR: Diagnostic odds ratios; LR: Likelihood ratios; S: Advantage index, a ranking metric for diagnostic performance; higher values indicate superior overall accuracy; RR_Se/RR_Sp vs CL: Relative ratio compared to cervical length (reference); CL: Cervical length; ACA: Anterior cervical angle; IOS: Internal os strain; EOS: External os strain; ECI: Elasticity contrast index; HR: Hardness ratio.

For individual E-cervix parameters, ECI, IOS, EOS, and IOS/EOS ratio showed DORs broadly similar to CL (4.00-5.47). HR demonstrated the strongest single-parameter discrimination [area under the curve (AUC) = 0.80, 95%CrI: 0.72-0.85; sensitivity = 0.76; specificity = 0.70; DOR = 7.71]. IOS showed a distinct profile with lower sensitivity (0.56) but higher specificity (0.80), suggesting its utility for predicting PTB. Other parameters, including ECI, EOS, and the IOS/EOS ratio, yielded DORs (range: 4.05-5.47) that were slightly higher than or comparable to CL.

The combined multi-parameter model achieved the highest overall diagnostic performance (AUC = 0.91, 95%CrI: 0.83-0.95; sensitivity = 0.83, 95%CrI: 0.74-0.89; specificity = 0.83, 95%CrI: 0.73-0.90; DOR = 26.23, 95%CrI: 10.47-54.64), outperforming all individual indices. These findings were consistent with the summary receiver operating characteristic curves (Figure 3), where the combined model was positioned closest to the optimal upper-left corner, followed by ACA and HR, whereas CL was furthest from the optimal region.

Figure 3 Summary receiver operating characteristic curves for preterm birth prediction. Combined model (brown) shows the highest diagnostic accuracy; cervical length (red) is the lowest. Axes represent sensitivity and 1-specificity. SROC: Summary receiver operating characteristic; TPR: True positive rate; FPR: False positive rate; CL: Cervical length; ACA: Anterior cervical angle; ECI: Elasticity contrast index; IOS: Internal os strain; EOS: External os strain; HR: Hardness ratio.

Heterogeneity and subgroup analysis

No threshold effects were detected for any parameter (Spearman P > 0.05). Substantial non-threshold heterogeneity was observed for CL (I² = 76.10%), ECI (I² = 70.50%), IOS/EOS ratio (I² = 84.80%), and the combined model (I² = 85.0%), whereas IOS and EOS showed minimal heterogeneity (I² = 0.00% for both). Therefore, subsequent pooling was conducted under a random-effects framework (Table 3).

Table 3 Summary of heterogeneity analysis between E-cervix parameters and conventional measurement parameters.

Parameters	Studies (n)	Pooled DOR (95%CI)	Cochran Q	P value (Q test)	I2 (%)	Spearman r	P value (Spearman)
CL	10	3.68 (1.88-7.22)	37.66	0.00	0.76	0.26	0.47
ACA	4	18.02 (7.92-40.98)	5.10	0.16	0.41	-0.80	0.20
ECI	9	4.92 (2.62-9.23)	23.75	0.00	0.71	-0.23	0.59
IOS	8	5.30 (3.88-7.25)	5.83	0.56	0.00	0.68	0.06
EOS	6	4.00 (2.81-5.69)	3.54	0.62	0.00	0.60	0.21
IOS/EOS	6	3.91 (1.45-10.57	32.90	0.00	0.85	-0.14	0.79
HR	11	10.92 (6.66-17.89)	20.32	0.03	0.51	0.45	0.16
Combined	7	35.27 (9.75-127.55)	40.02	0.00	0.85	-0.25	0.59

DOR: Diagnostic odds ratios; CI: Confidence interval; CL: Cervical length; ACA: Anterior cervical angle; IOS: Internal os strain; EOS: External os strain; ECI: Elasticity contrast index; HR: Hardness ratio.

Meta-regression incorporating pregnancy status, gestational age, risk stratification, study design, and sample size (Supplementary Table 2) showed no significant associations for most indices (all P > 0.05) (Supplementary Table 2). For most individual indices, including CL, ACA, IOS, EOS, and HR, no significant associations were observed (all P > 0.05). Regarding the combined model, pregnancy status significantly modified DOR (0.16, 95% confidence interval: 0.03-0.75; P = 0.03). Gestational age (P = 0.07) and study design (P = 0.06) were marginally associated with the performance of the combined model. These marginal results should be interpreted with caution due to the relatively small number of included studies and the wide confidence intervals observed in certain parameters (e.g., ACA), which may suggest limited statistical power for these specific covariates.

Subgroup analyses (Supplementary Table 3) indicated stable performance for CL across strata (sensitivity = 0.59-0.67; specificity = 0.59-0.74; all P ≥ 0.05). ACA remained highly accurate but showed evidence of small-sample optimism (P < 0.05). Among E-cervix parameters, HR showed the most robust single-parameter performance, with better results in low-risk populations and larger cohorts (P < 0.05). The combined multi-parametric model consistently delivered the superior diagnostic performance across nearly all subgroups, particularly in low-risk populations (sensitivity = 0.91, specificity = 0.88). Although its specificity was moderated by pregnancy status (P = 0.01) and sample size (P = 0.03), the combined model outperformed all individual parameters regardless of study scale (e.g., sensitivity = 0.98 in small cohorts vs 0.80 in large cohorts). Even in high-risk cohorts and larger-scale studies (n ≥ 100 subjects: Sensitivity = 0.80, specificity = 0.83), it demonstrates its broad clinical applicability across diverse settings.

Sensitivity analysis and publication bias

Sensitivity analysis using Stata (midas) yielded frequentist estimates consistent with the primary Bayesian results, with an identical ranking: The combined model and HR performed best, followed by ACA, while CL showed moderate accuracy (Supplementary Table 4). Deeks’ funnel plot indicated no significant asymmetry for most outcomes (all P > 0.05), but suggested possible small-study effects/publication bias for ECI (P = 0.02) (Supplementary Figure 1).

DISCUSSION

Summary of findings and comparative advantages

E-cervix elastography provides a functional assessment of cervical biomechanics and may detect cervical remodeling earlier than conventional morphological measures do[29]. In this systematic review and head-to-head network meta-analysis of 15 studies involving 1965 pregnancies, E-cervix-derived indices showed a clear diagnostic advantage over CL for predicting PTB. Notably, the multi-parametric combined model achieved the highest overall accuracy, while the HR emerged as the most robust single biomechanical parameter.

Compared to previous meta-analyses[30,31], our study offers two primary advantages. First, to our knowledge, it is the first network meta-analysis to provide a hierarchical ranking of E-cervix indices and conventional markers, facilitating indirect comparisons where head-to-head data were lacking. Second, through rigorous meta-regression and subgroup analyses, we isolated the impact of pregnancy status and risk stratification. This granular evidence provides a more precise framework for personalized clinical decision-making than earlier reviews that pooled heterogeneous populations.

Diagnostic value of E-cervix vs CL

Although CL remains the guideline-recommended marker[32], its diagnostic performance in our analysis was only moderate, characterized by a high false-negative rate, a finding consistent with previous reports[33]. In contrast, E-cervix parameters, especially HR, demonstrated a significantly higher advantage index and superior discriminatory power.

This shift from morphological to functional assessment is biologically plausible. By quantifying cervical strain and stiffness, E-cervix can capture early biomechanical changes such as collagen rearrangement and softening, whereas CL reflects only geometric shortening of the cervix and typically becomes abnormal only after overt structural changes have occurred[34,35]. Thus, during the early phase of cervical functional decompensation, CL may still fall within the “normal” range, while the E-cervix has already detected an elevated risk[36]. Our findings support E-cervix not as a replacement, but as a crucial functional complement to CL, particularly in women with borderline CL values or additional risk factors, in order to improve both the sensitivity and precision of risk stratification.

Effectiveness and nature of combined diagnostic models

The multi-parametric combined model demonstrated the most favorable diagnostic profile, achieving the highest advantage index (S: 12.90) and AUC (0.91). By integrating E-cervix indices with morphological markers (CL or ACA), these studies established a hybrid morphological biomechanical framework. This approach capitalizes on the synergy between structural and functional data: Whereas CL and ACA assess macro-anatomical configuration, E-cervix evaluates microstructural integrity[18,29].

Notably, the combined model maintained high diagnostic accuracy in low-risk populations (sensitivity = 0.91, specificity = 0.88), a setting in which isolated CL measurement often shows limited sensitivity because cervical shortening is relatively uncommon. Consequently, incorporating biomechanical markers into the assessment may refine clinical decision-making regarding prophylactic interventions (e.g., progesterone therapy) by more accurately distinguishing truly high-risk pregnancies from those with isolated morphological alterations but preserved cervical tissue competence[37].

Performance of ACA and other single parameters

Among conventional morphological indices, ACA exhibited substantially better discriminatory ability than CL (with a clearly higher DOR), suggesting that geometric features such as cervical funneling and open-angle configuration play an important role in PTB risk assessment[38]. However, only four ACA-related studies were included, with differences in measurement techniques, cut-off definitions and study populations; its high ranking in the network analysis therefore requires confirmation in larger, prospective cohorts with standardized imaging protocols.

Apart from HR, other single E-cervix parameters showed more heterogeneous performance. The overall diagnostic efficacy of ECI, IOS, EOS and the IOS/EOS ratio was generally no worse than that of CL, with DORs around 4-5. Interestingly, the IOS exhibited a ‘rule-in’ diagnostic profile, characterized by high specificity despite modest sensitivity. This suggests that different E-cervix parameters could be strategically positioned within a stepwise screening pathway, for instance, using HR for initial screening and IOS to confirm high-risk status[31].

Heterogeneity, pregnancy status, and bias

Substantial non-threshold heterogeneity was observed for CL (I² = 76.1%) and the combined model (I² = 85.0%), likely driven by variations in patient populations and gestational ages. For the combined model, this variance largely stems from inconsistent definitions and constructions of “multi-parametric” frameworks across the included studies. These approaches ranged from relatively simple dual-index combinations to comprehensive structures - such as Ma et al’s six-parameter model[27] (ACA, ECI, HR, IOS, EOS, and IOS/EOS ratio) compared with Kwon et al’s dual-index approach[28] (CL and IOS/EOS ratio).

An examination of study-level outcomes reveals substantial variability, where diagnostic accuracy appears to scale with increasing model complexity. Specifically, reported AUC values range from approximately 0.71 in simpler dual-parameter models to over 0.94 in more complex multi-index frameworks. This trend suggests that higher model complexity - often a result of study-specific optimization - may artificially inflate apparent performance. Consequently, pooling such heterogeneously composed models risks introducing structural aggregation bias, wherein the most optimized or parameter-rich models disproportionately influence the final estimate. Thus, the summary effect (AUC = 0.91) should be interpreted as a reflection of the synergistic potential of hybrid morphological-biomechanical integration, rather than a validated performance benchmark for any single standardized clinical protocol.

Although pregnancy status (singleton vs multiple) significantly modified performance (P = 0.03), it failed to fully explain the variance, reinforcing parameter selection as a primary source of statistical uncertainty[39]. Finally, the publication bias detected for ECI (P = 0.02) necessitates cautious interpretation and further validation in larger, pre-specified cohorts.

Limitations

Several limitations of this study warrant consideration. First, most included studies were single-center observational cohorts conducted in China, and the E-cervix parameters are currently vendor-specific, which collectively may limit the global generalizability of the findings. Second, the quality of evidence was not formally assessed using the Grading of Recommendations, Assessment, Development and Evaluations framework, which should be considered when interpreting the strength of our conclusions. Third, cut-off values in many included studies were derived post-hoc rather than pre-specified; this retrospective optimization of thresholds may have inflated the diagnostic accuracy estimates and introduced potential selection bias.

Furthermore, significant heterogeneity existed in the definition and components of the combined multi-parameter models across different studies. Pooling these heterogeneous composite models may have led to partial overestimation of their diagnostic efficacy, particularly when more complex or internally optimized models were included. This may impact the stability, generalizability, and external reproducibility of the pooled results for the combined approach. Lastly, the lack of standardized acquisition protocols remains a technical hurdle. Future large-scale, multicenter prospective trials with standardized measurement techniques and universal thresholds are essential to validate and integrate these biomechanical markers into routine clinical practice for precise PTB prediction.

CONCLUSION

Cervical elastography provides higher diagnostic value than conventional CL measurement for predicting PTB. Among individual indices, HR shows the most favorable performance, while multi-parametric morphological-biomechanical combined models achieve the best overall accuracy. Nevertheless, due to the limited number and inherent heterogeneity of the studies included in this meta-analysis, these conclusions should be interpreted with caution and require further validation through more high-quality prospective investigations.

ACKNOWLEDGEMENTS

We thank all investigators of the included primary studies for making their data available.