Multi-dimensional predictive model for diminished ovarian reserve in Hashimoto's thyroiditis: development and application.

Both intra- and inter-observer reproducibility for bilateral OV, AFC, and blood flow parameters were excellent, with all ICC values exceeding 0.75. Intra-observer consistency was higher than inter-observer consistency (Table  1 ). Table 1 Intra- and inter-observer consistency for ultrasound measurements Intra-observer ICC (95% CI) Inter-observer ICC (95% CI) Bilateral ovarian volume 0.997 0.979 (95% CI = 0.993–0.999) (95% CI = 0.947–0.992) Total bilateral AFC 0.988 0.957 (95% CI = 0.970–0.995) (95% CI = 0.894–0.983) VI 0.997 0.989 (95% CI = 0.993–0.999) (95% CI = 0.993–0.999) FI 0.904 0.896 (95% CI = 0.774–0.961) (95% CI = 0.738–0.959) VFI 0.997 0.989 (95% CI = 0.993–0.999) (95% CI = 0.973–0.996) Intra- and inter-observer consistency for ultrasound measurements Among the 300 participants, 162 (54.0%) were diagnosed with DOR, while 138 (46.0%) were not. No statistically significant differences in baseline characteristics were observed between the training and validation sets (Table  2 ). Table 2 Baseline characteristics of participants in the training and validation sets Variable Group P -value Training set ( n  = 210) Validation set ( n  = 90) Age [years] 33.88 ± 4.68 34.41 ± 4.48 0.749 Body mass index [kg/m²] 23.73 ± 1.87 23.77 ± 2.11 0.182 Age at menarche [years] 13.12 ± 0.94 13.31 ± 0.92 0.649 Menstrual cycle [days] 29.87 ± 2.81 29.90 ± 2.81 0.885 HT subtype, n (%) 0.217  -Euthyroid 153(73.0%) 74(82.0%)  -Hypothyroidism 30(14.2%) 9(10.0%)  -Hyperthyroidism 27(12.8%) 7(8.0%) TPOAb positive, n (%) 172(81.9%) 74(82.2%) TGAb positive, n (%) 150(71.4%) 68(75.5%) 0.199 Mean ovarian volume [cm³] 3.32 ± 1.27 3.18 ± 1.21 0.798 Total bilateral AFC [count] 12.82 ± 6.71 11.78 ± 6.33 0.807 History of pelvic inflammation, n (%) 0.052  -Yes 120(57.1%) 62(68.9%)  -No 90(42.8%) 28(31.1%) History of gynecological surgery, n (%) 0.855  -Yes 133(63.3%) 58(64.4%)  -No 77(36.7%) 32(35.6%) History of endometriosis, n (%) 0.319  -Yes 103(49.1%) 50(55%)  -No 107(50.9%) 40(45%) Fertility history, n (%) 0.157  -Yes 144(68.6%) 69(76.7%)  -No 66(31.4%) 21(23.3%) Ovarian VI 3.84 ± 4.86 4.13 ± 6.15 0.141 Ovarian FI 32.18 ± 19.5 32.584 ± 24.3 0.654 VFI 1.32 ± 2.27 1.50 ± 2.82 0.183 Follicle-stimulating hormone [mIU/ml] 7.89 ± 3.49 7.79 ± 3.64 0.740 Luteinizing hormone [mIU/ml] 6.58 ± 3.91 6.53 ± 3.43 0.421 Testosterone [ng/ml] 0.387 ± 0.19 0.408 ± 0.17 0.388 Baseline characteristics of participants in the training and validation sets History of endometriosis, n (%) LASSO regression identified four independent predictors of DOR: TGAb positivity, VI, bilateral AFC, and a history of pelvic inflammation (Fig.  4 ). Fig. 4 LASSO regression coefficient path diagram ( A ) and cross-validation curve ( B ) for the training set LASSO regression coefficient path diagram ( A ) and cross-validation curve ( B ) for the training set Based on these predictors, a nomogram was constructed using the rms package in R software (Fig.  5 ). The nomogram calculates the individual probability of DOR by summing the scores corresponding to each predictor. Fig. 5 Nomogram for predicting DOR in reproductive-aged women with HT based on transvaginal three-dimensional ultrasound. (The red arrow indicates a reproductive-aged HT woman with a history of pelvic inflammation, ovarian VI = 2.3, and 7 bilateral antral follicles. Her total score is 352, and the model estimates her probability of DOR as 0.992.) Nomogram for predicting DOR in reproductive-aged women with HT based on transvaginal three-dimensional ultrasound. (The red arrow indicates a reproductive-aged HT woman with a history of pelvic inflammation, ovarian VI = 2.3, and 7 bilateral antral follicles. Her total score is 352, and the model estimates her probability of DOR as 0.992.) ROC curve analysis demonstrated that the nomogram had an AUC of 0.989 (95% CI: 0.981–0.998). Delong’s test confirmed that the AUC of the nomogram was significantly greater than that of each individual risk factor (all p  < 0.05) (Fig.  6 ). Fig. 6 ROC curves comparing the nomogram model with individual independent risk factors (history of pelvic inflammation, VI, bilateral AFC, and TGAb positivity) ROC curves comparing the nomogram model with individual independent risk factors (history of pelvic inflammation, VI, bilateral AFC, and TGAb positivity) In the training and validation sets, the nomogram yielded AUC values of 0.989 (95% CI: 0.981–0.998) and 0.985 (95% CI: 0.968–1.000), respectively. The corresponding sensitivities and specificities were 0.964/0.930 for the training set and 0.882/0.974 for the validation set (Fig. 7 A; Table 3 ). To more objectively and realistically assess the internal generalizability of this nomogram model, internal validation was performed on the training set using the bootstrap resampling method ( n  = 1000). The results showed that the model’s predictive performance remained consistently high across most resampled datasets, suggesting good discriminative ability and result stability (Fig. 7 B). In addition, although some variability in the ROC curves was observed among different bootstrap samples, the overall range of variation was small. This indicates that the model performance is unlikely to be attributable solely to a particular single data split and instead demonstrates a degree of robustness. Fig. 7 A ROC curves of the nomogram model for predicting DOR in the training set ( n  = 210) and validating set ( n  = 90); B ROC curves for bootstrap internal validation ( n  = 1000) A ROC curves of the nomogram model for predicting DOR in the training set ( n  = 210) and validating set ( n  = 90); B ROC curves for bootstrap internal validation ( n  = 1000) Table 3 Accuracy of the nomogram in estimating the risk of DOR Variable Training cohort Validating cohort Area under the ROC curve 0.989 (0.981–0.998) 0.985 (0.968-1.000) Cut-off score 0.452 0.780 Youden index 0.894 0.885 Sensitivity, % 0.964 0.882 Specificity, % 0.930 0.974 Positive predictive value, % 0.938 0.978 Negative predictive value, % 0.959 0.864 Positive likelihood ratio 13.766 34.412 Negative likelihood ratio 0.039 0.121 Accuracy of the nomogram in estimating the risk of DOR Calibration curves showed strong concordance between predicted and observed probabilities of DOR in both the training and validation sets (Fig.  8 ). Fig. 8 Calibration curves of the nomogram model for the training set ( A ) and validation set ( B ). (The red line represents the ideal prediction; the grey line represents the model performance; the shaded area represents the 95% confidence interval.) Calibration curves of the nomogram model for the training set ( A ) and validation set ( B ). (The red line represents the ideal prediction; the grey line represents the model performance; the shaded area represents the 95% confidence interval.) DCA demonstrated a high net clinical benefit when the threshold probability exceeded 1% in the training set and 3% in the validation set (Fig.  9 ). Fig. 9 Clinical decision curves of the nomogram model for the training set ( A ) and validation set ( B ) Clinical decision curves of the nomogram model for the training set ( A ) and validation set ( B ) At threshold probabilities exceeding 60%, the number of high-risk patients identified by the model closely approximated the actual number of cases, confirming its substantial clinical net benefit (Fig.  10 ). Fig. 10 Clinical impact curve of the nomogram model. (The red curve represents the number of high-risk patients identified by the model; the blue dashed line represents the actual number of high-risk patients at each threshold probability.) Clinical impact curve of the nomogram model. (The red curve represents the number of high-risk patients identified by the model; the blue dashed line represents the actual number of high-risk patients at each threshold probability.)

A prospective cross-sectional study was conducted, enrolling 328 reproductive-aged women diagnosed with HT who attended the Second Affiliated Hospital of Fujian Medical University between February 2024 and June 2025. Following the strict application of inclusion and exclusion criteria, 300 eligible participants were included and randomly assigned to a training set (210 cases) and a validation set (90 cases) in a 7:3 ratio using a random number table (Fig.  1 ). The study protocol was approved by the Medical Ethics Committee of the Second Affiliated Hospital of Fujian Medical University (Approval No. 2024 − 112), and written informed consent was obtained from all participants. Fig. 1 The flowchart of included and divided patients into the training set and validating set in this study The flowchart of included and divided patients into the training set and validating set in this study Inclusion criteria: (1) Married women aged 25–45 years of reproductive age; (2) Newly diagnosed HT with no prior treatment or untreated HT with no history of levothyroxine replacement therapy; (3) Intact bilateral ovaries with no history of ovarian surgery. Exclusion Criteria: (1) Bilateral or unilateral oophorectomy; (2) Recent use of sex hormones; (3) Bilateral tubal obstruction or ovulation disorders; (4) Ovarian hypofunction or premature ovarian failure secondary to bilateral or unilateral ovarian surgery; (5) Bilateral or unilateral salpingectomy; (6) Presence of other autoimmune diseases known to affect ovarian function; (7) Other ovarian pathologies (e.g., endometriotic cysts, tumours). HT diagnosis is based on the presence of the following: diffuse goitre with a firm texture (particularly with pyramidal lobe enlargement); ultrasonographic findings of heterogeneous thyroid parenchyma with diffuse hypoechogenicity and short linear hyperechoic strands forming a reticular pattern; and significantly elevated serum levels of TPOAb and TGAb [ 14 ]. Diagnosis of DOR: Based on the Expert Consensus on Clinical Application of Anti-Müllerian Hormone (2023 Version) [ 15 ] and the Expert Consensus on Clinical Diagnosis and Treatment of Diminished Ovarian Reserve (2022 Version) [ 16 ], DOR is defined as an AMH level < 1.1 ng/mL in women aged ≥ 40 years or < 1.68 ng/mL in women aged < 40 years. A GE Voluson E10 colour Doppler ultrasound system equipped with a 3D endovaginal volume probe (frequency range: 5–9 MHz) was utilised. Transvaginal three-dimensional ultrasound was performed between days 3 and 7 of the menstrual cycle by a sonographer with 5 years of experience. Participants were instructed to empty their bladders and adopt the lithotomy position. The ultrasound probe, covered with a condom, was inserted into the vaginal fornix. Routine two-dimensional scanning was initially conducted to evaluate ovarian size, morphology, and echogenicity. Upon identification of the maximum ovarian diameter in 2D mode, the 3D mode was activated, and SonoAVC automatic measurement software was employed to detect AFC. Follicles were listed according to diameter, colour-coded, and manually adjusted to correct any counting or measurement discrepancies. Power Doppler imaging was then overlaid, and 3D volume scanning was conducted. Following acquisition, transparent mode within Volume Analysis was used to manually outline the ovarian contour (rotation angle: 15°, 12 contours), allowing reconstruction of ovarian volume (OV). Blood flow parameters—including vascularization index (VI), flow index (FI), and vascularization flow index (VFI)—were quantified using Volume Histogram. For statistical analysis, the total AFC was calculated as the sum of bilateral counts, while OV, VI, FI, and VFI were computed as the bilateral averages (Figs.  2 and 3 ). Fig. 2 A , B Automated analysis of antral follicles in the right ovary using SonoAVC technology. The software individually color-codes each identified follicle and provides objective measurements of their mean and absolute diameters and volumes A , B Automated analysis of antral follicles in the right ovary using SonoAVC technology. The software individually color-codes each identified follicle and provides objective measurements of their mean and absolute diameters and volumes Fig. 3 Measurement of ovarian volume (OV), vascularization index (VI), flow index (FI), and vascularization flow index (VFI) using VOCAL technology Measurement of ovarian volume (OV), vascularization index (VI), flow index (FI), and vascularization flow index (VFI) using VOCAL technology Twenty healthy reproductive-aged women were recruited for reliability analysis. A double-blind method was adopted: two qualified sonographers independently repeated the measurements using the same equipment, and a third physician recorded the data. The first sonographer performed the initial measurements, followed by the second sonographer after a 10-minute interval. The first sonographer then repeated the measurements after another 10 min. Ultrasound parameters (OV, AFC, VI, FI, VFI) were recorded. Clinical data included age, body mass index (BMI), menstrual cycle regularity, age at menarche, history of pelvic inflammation, gynaecological surgery, endometriosis, miscarriage, and laboratory values (FSH, LH, testosterone, HT subtype, thyroid-stimulating hormone [TSH], T3, and T4). This study aims to build a predictive model for DOR. The overall incidence rate of all participants ultimately enrolled in the study is 0.540. According to the sample size calculation formula proposed by Riley et al. [ 17 ], we have calculated that our research requires a minimum of 282 participants to be enrolled. Statistical analysis was performed using SPSS 26.0, MedCalc, and R software. Intraclass correlation coefficients (ICC) and Bland-Altman plots were used to assess intra- and inter-observer consistency. The Shapiro-Wilk test and histograms were used to evaluate normality. Data conforming to a normal distribution were expressed as mean ± standard deviation and compared using the independent samples t-test. Non-normally distributed data were expressed as median (interquartile range) [M (Q1, Q3)] and analysed using the Mann-Whitney U test. Categorical variables were expressed as counts (percentages) and compared using the chi-square test. A 10-fold cross-validation curve of λ versus cross-validation error (mean squared error, MSE) was first plotted; the λ that minimized the MSE (λ_min) was identified, and λ_min was chosen to select independent predictors. Under this criterion, LASSO regression was conducted on the training set to identify independent risk factors for DOR and construct the nomogram. ROC curves were generated to determine optimal cut-off values based on the Youden index. Calibration curves, DCA, and clinical impact curves were employed to evaluate the model’s performance. A two-tailed P  < 0.05 was considered statistically significant.

In this study, 54.0% of reproductive-aged women with HT were diagnosed with DOR, further supporting the association between HT and reproductive impairment. By integrating transvaginal three-dimensional ultrasound parameters with clinical data, we developed a visual risk assessment tool that identified history of pelvic inflammation, VI, AFC, and TGAb positivity as key predictive factors. These variables reflect immune dysregulation (TGAb), inflammatory damage (pelvic inflammation), and ovarian functional status (VI, AFC), thereby providing a multidimensional understanding of the pathogenesis of DOR in HT. The association between HT and DOR is likely mediated by immune injury, an inflammatory microenvironment, and ovarian microcirculatory dysfunction. TGAb positivity, a significant risk factor, was present in 75.2% of HT patients in a meta-analysis [ 9 ], and negatively correlated with AMH levels. Mechanistically, TGAb may cross-react with the ovarian zona pellucida antigen ZP3 via molecular mimicry, activating complement pathways and inducing granulosa cell apoptosis and follicular atresia [ 18 , 19 ]. In our study, TGAb-positive patients exhibited a 32% reduction in bilateral AFC, supporting the role of autoimmune-mediated follicular damage. A history of pelvic inflammation demonstrated the highest standardised coefficient in the LASSO regression model, underscoring its critical influence. Regulatory T cell (Treg) dysfunction associated with HT may exacerbate ovarian cortical injury resulting from pelvic inflammation [ 20 ]. In HT patients with a history of pelvic inflammation, VI was reduced by 41%, suggesting that the combination of autoimmunity and chronic inflammation accelerates DOR progression through microcirculatory compromise. Chronic inflammation may trigger endothelial apoptosis and vascular remodelling, creating a vicious cycle with follicular atresia. These findings underscore VI as a potential marker for early ovarian reserve decline [ 21 , 22 ]. The integration of transvaginal three-dimensional ultrasound parameters represents a methodological advancement. The SonoAVC system reduces subjectivity in AFC measurements, with AFC ≤ 7 being associated with a significantly increased risk of DOR. AFC has been shown to correlate negatively with age and to predict ovarian response more effectively than AMH, particularly in younger women [ 23 , 24 ]. Moreover, VI values quantified using VOCAL technology demonstrated a negative correlation between ovarian microvascular density and autoantibody titres. The model exhibited excellent predictive performance, with AUCs of 0.989 and 0.985 in the training and validation sets, respectively, and sensitivities and specificities exceeding 88%. This supports its utility as an integrated tool for assessing immune, morphological, and functional aspects of ovarian reserve. AMH offers a static snapshot of follicular number yet remains blind to the dynamic choreography of ovarian microcirculation. In Hashimoto thyroiditis, persisting immune inflammation destabilises granulosa cell AMH release, yielding false-normal readings in fifteen to twenty per cent of women who already suffer diminished ovarian reserve. Our cohort mirrors this hazard, as twenty-three per cent of euthyroid AMH-normal patients still carried occult DOR risk, exposing the false-negative pitfall of solitary AMH appraisal. Conventional two-dimensional ultrasound compounds the problem by failing to capture ovarian volume and perfusion in the same acquisition frame. By simultaneously quantifying OV, VI, FI and VFI through three-dimensional VOCAL technology, we weave morphology, blood supply and follicular competence into a single continuous assessment chain, overcoming both the quantitative-only vision of AMH and the micro-environmental blindness inherent in planar imaging. Clinically, the model facilitates precise risk stratification. It demonstrated a net benefit when threshold probabilities exceeded 3%, allowing for earlier identification of DOR risk compared to traditional markers. At threshold probabilities above 60%, it accurately identified high-risk individuals, supporting timely fertility preservation strategies. Notably, the model identified 23% of HT patients with normal AMH levels but latent risk of DOR, underscoring its capacity to overcome the limitations of AMH in autoimmune contexts. Subgroup analyses conducted on HT patients with normal AMH levels ( n  = 87) revealed that 20 patients (20/87, 23.0%) were assessed as high risk for DOR using the nomogram, defined as latent DOR (risk probability ≥ 60%). The core characteristics of this subgroup included: a mean bilateral AFC of 9.3 ± 2.1, which was significantly lower than that of both the normal AMH and low-risk nomogram groups (15.6 ± 3.4, P  < 0.001). The mean VI was 3.2 ± 0.8, compared to 5.7 ± 1.2 in the low-risk group ( P  < 0.001). Furthermore, the positive rate of TGAb) was significantly higher in the latent DOR group (85.0%) compared to the low-risk group (42.3%, P  < 0.001). After a follow-up period of three months, among the 20 patients identified with latent DOR, 4 (20.0%) experienced a decline in AMH to below the normal range, while only 2 patients (3.0%) in the low-risk group exhibited a decrease in AMH levels. This suggests that latent DOR carries a risk of progression to overt DOR, thereby validating the nomogram’s effectiveness in identifying potential risks. It is noteworthy that, in the present study, thyroid-related indicators—including TSH, FT3, FT4, and TPOAb—were excluded following LASSO regression analysis. This finding may be attributed to the primary objective of LASSO, which is to enhance model parsimony and generalizability rather than to establish causal relationships. When variables such as AFC, VI, TGAb, and a history of pelvic inflammation already provide strong predictive information, thyroid function parameters may be shrunk to zero coefficients due to their limited incremental predictive value. In addition, the relatively small sample size of the thyroid dysfunction subgroups (with hypothyroidism and hyperthyroidism comprising 30 and 27 cases, respectively, in the training set, and 9 and 7 cases in the validation set) may have resulted in insufficient statistical power, leading to instability and subsequent exclusion during variable selection. Therefore, within this dataset and modeling framework, TSH, FT3, FT4, and TPOAb did not demonstrate independent incremental predictive value. However, this does not imply a lack of association with DOR, nor does it exclude the potential roles of untreated thyroid dysfunction as confounding, effect-modifying, or mediating factors in the development of DOR. In daily reproductive medicine practice, the nomogram can be applied across key scenarios to enhance HT patients’ reproductive health management. For screening, it targets newly diagnosed HT women (25–45 years, with fertility needs), prioritizing those with pelvic inflammation history or high TGAb titers, enabling rapid DOR risk assessment via 3D ultrasound and TGAb testing. In follow-ups, it guides stratified monitoring: low-risk patients (≤ 30% probability) yearly, medium-risk (30%-60%) semi-annually, high-risk (≥ 60%) quarterly, and tracks treatment efficacy. For fertility counseling, it aids personalized advice—alerting AMH-normal but high-risk patients to early fertility efforts and guiding pre-ART pretreatment for high-risk cases, boosting clinical translation value.

This study has several limitations. First, only internal validation was performed without multicenter external validation, potentially overestimating the model’s performance or even overfitting. Second, the single-center design and recruitment of participants from a specific clinical population introduce health-seeking bias, with exclusion of unmarried individuals and those with a history of ovarian surgery, which may compromise the external validity of the results. Third, the 3-month follow-up is likely insufficient, as AMH may exhibit short-term variability rather than true decline. Thus, attributing the observed changes to diminished ovarian reserve may be premature. Cautious interpretation is warranted, and longer-term studies are needed for confirmation. Fourth, although patients receiving levothyroxine were excluded, the specific impacts of thyroid hormone levels were not thoroughly analyzed, and subclinical thyroid dysfunction may exert confounding effects, affecting the accuracy of the association strength between risk factors and DOR.

Hashimoto’s thyroiditis (HT), also known as autoimmune thyroiditis (AIT), is a common organ-specific autoimmune disorder. A characteristic feature of HT is the presence of thyroid peroxidase antibody (TPOAb) and thyroid globulin antibody (TGAb), with approximately 95% of patients testing positive for TPOAb and 60–80% for TGAb [ 1 , 2 ]. Although its pathogenesis involves genetic, environmental, and epigenetic factors, it remains incompletely elucidated. In recent years, increasing attention has been directed toward the association between HT and reproductive failure in women of reproductive age. A growing body of evidence suggests that HT is the most prevalent autoimmune disorder linked to premature ovarian insufficiency (POI) [ 3 – 7 ]. Chen et al. [ 8 ] reported a higher susceptibility to POI among women with HT, while a meta-analysis by Li et al. [ 9 ] confirmed a significantly increased risk of diminished ovarian reserve in reproductive-aged women with HT. Ovarian reserve refers to the ability of the ovarian cortex to generate and mature follicles and is directly associated with female reproductive potential [ 10 ]. Diminished ovarian reserve (DOR) is defined by a reduction in both the quantity and quality of remaining oocytes, which significantly compromises the success rates of assisted reproductive technology (ART) [ 11 ]. Current assessments of ovarian reserve primarily rely on biochemical markers—such as follicle-stimulating hormone (FSH), luteinizing hormone (LH), estradiol (E2), inhibin B, and anti-Müllerian hormone (AMH)—as well as ultrasound parameters such as antral follicle count (AFC). Among these, AMH is regarded as a relatively reliable marker due to its stability across the menstrual cycle and limited susceptibility to external influences such as body weight and lifestyle factors [ 10 ]. However, the clinical evaluation of ovarian reserve in HT patients faces two notable challenges. First, although serum AMH is considered the gold standard, it lacks the ability to dynamically reflect fluctuations in ovarian function and may be biased under autoimmune conditions. Second, AFC measurements using two-dimensional ultrasound are highly operator-dependent and prone to counting errors, limiting their utility in precision medicine. Our preliminary studies revealed strong correlations between transvaginal three-dimensional ultrasound parameters—AFC, ovarian volume, and VI—and AMH in patients with gynaecological disorders ( r  = 0.73–0.80, P  < 0.001). The SonoAVC system applied in this study reduced AFC measurement error to below 8% through multicolour coding and automated algorithms [ 12 ]. Additionally, the use of three-dimensional power Doppler ultrasound combined with VOCAL software facilitated innovative, quantitative evaluation of ovarian microvascular perfusion [ 13 ], thereby providing a solid methodological basis for further research. Although the association between HT and compromised ovarian reserve has been established, limitations in existing assessment technologies and the lack of comprehensive predictive models integrating transvaginal three-dimensional ultrasound with clinical data underscore the need for further investigation. This study addresses these gaps by constructing a predictive model that integrates immune indicators, morphological parameters, and blood perfusion indices. The objective is to enable early detection of ovarian functional decline in HT patients and to furnish visualised, evidence-based guidance for personalised reproductive health management, thereby enhancing fertility outcomes.