Inflammation at the crossroads of reproduction: SIRI as a prognostic signature of female infertility in hybrid regression-machine learning models.

This study establishes the Systemic Inflammatory Response Index (SIRI) as an independent, dose-dependent predictor of female infertility, with higher levels significantly increasing the risk. By combining regression and machine learning approaches, the findings underscore SIRI as a robust, clinically actionable biomarker, supporting individualized reproductive risk assessment, informing early intervention strategies, and facilitating personalized management.

To improve the robustness and reliability of our predictive framework, we employed a multi-dimensional feature selection strategy. Initially, feature importance was assessed using three distinct algorithms: Joint Mutual Information Maximization (JMIM), Logistic Regression (LR), and Random Forest (RF). Each method offered a unique perspective on variable relevance, enabling a comprehensive evaluation of the predictive value of each feature. The resulting rankings are presented in Fig.  4 A–C. Based on this analysis; we identified a core set of 16 primary predictors encompassing socioeconomic, clinical, anthropometric, and behavioral domains. These included the Systemic Inflammation Response Index (SIRI), age, body mass index (BMI), waist circumference, diabetes mellitus (DM), chronic kidney disease (CKD), chronic obstructive pulmonary disease (COPD), food security status, smoking status, hypertension, alcohol consumption, hyperlipidemia, triglyceride levels, total cholesterol, Homeostatic Model Assessment for Insulin Resistance (HOMA-IR), and anemia. Additionally, three context-dependent variables—ethnicity, educational attainment, and marital status—were retained for further investigation due to their potential, albeit less consistent, predictive contributions. These ranked features informed the development of five predictive models, which were systematically evaluated using the area under the curve (AUC) to assess discriminative performance. Among these, Model B (illustrated in Fig.  4 D) demonstrated the highest predictive accuracy, achieving the highest AUC score. This optimal model incorporated eight key predictors: SIRI, age, race, educational level, BMI, hypertension, diabetes status, and total cholesterol. Fig. 4 Feature selection and model performance comparison. A Ranking of feature importance using Random Forest; B Feature selection based on Joint Mutual Information Maximisation (JMIM); C Feature selection via Lasso Regression; D Comparison of AUC values across predictive models constructed using various feature combinations by radar chart (tenfold cross-validation) Feature selection and model performance comparison. A Ranking of feature importance using Random Forest; B Feature selection based on Joint Mutual Information Maximisation (JMIM); C Feature selection via Lasso Regression; D Comparison of AUC values across predictive models constructed using various feature combinations by radar chart (tenfold cross-validation) To enhance model performance and ensure generalizability, we implemented a nested cross-validation strategy for hyperparameter optimization. This rigorous approach mitigates overfitting and provides more reliable performance estimates. Specifically, the inner loop utilized tenfold cross-validation to fine-tune hyperparameters, while the outer loop employed fivefold cross-validation to evaluate model performance on independent subsets of data. This comprehensive resampling procedure led to the training and evaluation of approximately 16,262 models, underscoring the thoroughness of the optimization process. In the final modeling phase, we evaluated the predictive performance of eight machine learning algorithms: Logistic Regression (LR), Decision Tree Classifier (DTC), Support Vector Machine (SVM), Kernel SVM (KSC), k-Nearest Neighbors (KNN), Random Forest (RF), Neural Network (NNet), and eXtreme Gradient Boosting (XGBoost). Each algorithm was systematically assessed on both training and validation datasets to ensure robustness. Comparative results are presented in Fig.  5 and Table  4 . Among the tested models, XGBoost achieved the highest area under the curve (AUC) score (0.866), indicating superior predictive performance. Beyond its strong discriminative power, XGBoost demonstrated practical and clinical relevance, as illustrated in Fig.  5 C. Its selection as the final model was driven by its capacity to handle heterogeneous data types and its robustness in capturing complex, non-linear relationships inherent in the dataset. Fig. 5 ROC and DCA curves for each method. A ROC in the training set. B ROC in the validation set. C DCA curves in the training set. D DCA curves in the validation set Table 4 Evaluation metrics of the models constructed by each method Train metrics: Model AUC Accuracy (ACC) PPV (Precision) NPV Sensitivity (SEN) Specificity (SPE) F1 score MCC Kappa Brier score LR 0.634 0.599 0.837 0.157 0.599 0.600 0.674 0.124 0.088 0.236 DTC 0.656 0.796 0.816 0.164 0.796 0.870 0.806 0.068 0.068 0.168 SVM 0.631 0.565 0.839 0.152 0.565 0.555 0.644 0.123 0.081 0.236 KSC 0.656 0.568 0.844 0.157 0.568 0.556 0.647 0.138 0.091 0.220 KNN 0.834 0.690 0.885 0.237 0.690 0.675 0.747 0.313 0.235 0.186 RF 0.713 0.780 0.836 0.214 0.780 0.832 0.804 0.159 0.151 0.173 NNet 0.809 0.779 0.871 0.279 0.779 0.798 0.812 0.307 0.274 0.144 XGBoost 0.852 0.905 0.903 0.877 0.905 0.997 0.875 0.361 0.258 0.078 Validation metrics: Model AUC Accuracy (ACC) PPV (Precision) NPV Sensitivity (SEN) Specificity (SPE) F1 score MCC Kappa Brier score LR 0.613 0.590 0.834 0.149 0.590 0.592 0.667 0.109 0.076 0.240 DTC 0.686 0.817 0.828 0.211 0.817 0.889 0.822 0.122 0.121 0.157 SVM 0.609 0.551 0.835 0.144 0.551 0.542 0.633 0.102 0.066 0.241 KSC 0.655 0.573 0.856 0.167 0.573 0.554 0.651 0.172 0.112 0.219 KNN 0.844 0.711 0.887 0.248 0.711 0.700 0.764 0.328 0.254 0.176 RF 0.714 0.778 0.832 0.200 0.778 0.832 0.801 0.138 0.131 0.175 NNet 0.770 0.771 0.856 0.245 0.771 0.802 0.804 0.238 0.216 0.149 XGBoost 0.866 0.901 0.890 0.778 0.901 0.995 0.869 0.303 0.211 0.077 LR Logistic Regression, DTC Decision Tree Classifier, SVM Support Vector Machine, KSC Kernel SVM, KNN k-Nearest Neighbors, RF Random Forest, NNet Neural Network, XGBoost eXtreme Gradient Boosting. ROC and DCA curves for each method. A ROC in the training set. B ROC in the validation set. C DCA curves in the training set. D DCA curves in the validation set Evaluation metrics of the models constructed by each method LR Logistic Regression, DTC Decision Tree Classifier, SVM Support Vector Machine, KSC Kernel SVM, KNN k-Nearest Neighbors, RF Random Forest, NNet Neural Network, XGBoost eXtreme Gradient Boosting. To enhance the interpretability of the XGBoost model and construct the Dynamic Prediction Network (DPN), we employed SHapley Additive exPlanations (SHAP) to generate a comprehensive summary plot (Fig.  6 ). In this visualization, each point represents an individual patient, with the X-axis indicating the SHAP value, which reflects both the direction and magnitude of a specific feature’s influence on the model’s prediction. Features are ranked along the Y-axis, with those at the top contributing most substantially to the overall model performance. Fig. 6 Summary plot of SHAP values for the model constructed by XGBoost algorithm. The horizontal position “SHAP value” indicates whether the impact of the value is associated with a higher or lower prediction, and the color of each SHAP value point indicates whether the observed value is higher (red) or lower (blue). The vertical coordinates show the importance of the features, sorted by the importance of the variables in descending order, with the upper variables being more important to the model Summary plot of SHAP values for the model constructed by XGBoost algorithm. The horizontal position “SHAP value” indicates whether the impact of the value is associated with a higher or lower prediction, and the color of each SHAP value point indicates whether the observed value is higher (red) or lower (blue). The vertical coordinates show the importance of the features, sorted by the importance of the variables in descending order, with the upper variables being more important to the model To facilitate clinical application and support translational use, we developed a visual nomogram to estimate the probability of infertility based on these predictors. Additionally, we created an interactive web-based tool (Fig.  7 ), accessible at ( https://siri-machine-algorithm-femaleinfertility.streamlit.app/ ) which allows healthcare professionals to input patient-specific data and receive individualized infertility risk predictions with a single click on the “Predict” button. In the event of technical issues, users can restart the application by selecting “Quit” followed by “Reload.” This platform offers a practical and accessible solution for integrating predictive modeling into clinical decision-making at the point of care. Fig. 7 Online web calculator application based on XGBoost modeling Online web calculator application based on XGBoost modeling

A total of 3059 reproductive-aged women (18–45 years) from the NHANES 2015–2020 dataset were included in the present cross-sectional analysis. Among them, 339 individuals (11.1%) self-reported a history of infertility, while 2720 participants (88.9%) did not report infertility (Table  1 ). The mean age of participants was 31.5 ± 8.16 years, with women experiencing infertility being significantly older (34.3 ± 6.96 years) compared to their fertile counterparts (31.2 ± 8.23 years, p < 0.001). No statistically significant differences were observed across infertility status with respect to race/ethnicity (p = 0.627). However, non-Hispanic White women made up a larger proportion of the infertility group (33.0%) compared to other racial groups. Marital status showed a significant association with infertility: 70.5% of infertile women were married or living with a partner, compared to 48.8% of non-infertile women (p = 0.046). Education levels did not differ significantly between the two groups (p = 0.581), with the highest proportion of participants in both groups having some college or an associate degree. Smoking and alcohol behaviors revealed statistically significant group differences. A greater proportion of infertile women were former smokers (37.5% vs 25.1%, p = 0.007), although no significant differences were seen in current smoking behavior categories (p = 0.989). Alcohol consumption patterns differed significantly: 84.1% of infertile participants reported being current or former drinkers compared to 76.9% among the non-infertile (p = 0.004). Table 1 Baseline clinical characteristics by female infertility status: insights from NHANES Data (2015–2020) All Female infertility p-value (n = 3059) No (n = 2720) Yes (n = 339) Age (years) 31.5 ± 8.16 31.2 ± 8.23 34.3 ± 6.96 < 0.001 Race/ethnicity, n (%) 0.627  Mexican American 517 (16.9%) 455 (16.7%) 62 (18.3%)  Other Hispanic 360 (11.8%) 326 (12.0%) 34 (10.0%)  Non-Hispanic White 885 (28.9%) 773 (28.4%) 112 (33.0%)  Non-Hispanic Black 761 (24.9%) 686 (25.2%) 75 (22.1%)  Non-Hispanic Asian 373 (12.2%) 334 (12.3%) 39 (11.5%)  Other race—including multi-racial 163 (5.3%) 146 (5.4%) 17 (5.0%) Marital status (%) 0.046  Married/living with partner 1446 (51.42%) 1207 (48.8%) 239 (70.5%)  Widowed/divorced/separated and living alone 1366 (48.57%) 1266 (51.2%) 100 (29.5%) Education level, n (%) 0.581  < 9th grade 164 (5.9%) 151 (6.1%) 13 (3.9%)  9–11th grade 273 (9.8%) 238 (9.7%) 35 (10.5%)  High school graduate or equivalent 539 (19.3%) 474 (19.3%) 65 (19.6%)  Some college or AA degree 1042 (37.3%) 914 (37.2%) 128 (38.6%)  College graduate or above 773 (27.7%) 682 (27.7%) 91 (27.4%) Smoked at least 100 cigarettes, n (%) 0.007  Former smoker 811 (26.5%) 684 (25.1%) 127 (37.5%)  Never smoker 2246 (73.4%) 2034 (74.8%) 212 (62.5%) Current smoking behavior, n (%) 0.989  Current heavy smoker 380 (46.9%) 320 (46.8%) 60 (47.2%)  Current light smoker 107 (13.2%) 90 (13.2%) 17 (13.4%)  Occasional smoker 324 (40.0%) 274 (40.1%) 50 (39.4%) Alcohol consumption, n (%) 0.004  Former/current drinker 2187 (77.7%) 1902 (76.9%) 285 (84.1%)  Never drinker 626 (22.3%) 572 (23.1%) 54 (15.9%) BMI (kg/m 2 ) 29.37 ± 8.16 29.05 ± 8.05 31.72 ± 8.59 < 0.001 Waist circumference (cm) 95.9 ± 18.8 95.1 ± 18.50 102.2 ± 19.85 < 0.001 Hip circumference (cm) 109 ± 16.5 108.5 ± 16.2 113 ± 18.57 < 0.001 Fasting glucose (mg/dL) 101 ± 27 100 ± 25.2 107 ± 38.8 < 0.001 HOMA-IR 3.31 ± 4.43 3.21 ± 4.40 4.11 ± 4.55 0.092 Hypertension, n (%) 0.746  No 2877 (94.1%) 2560 (94.1%) 317 (93.5%)  Yes 182 (5.9%) 160 (5.9%) 22 (6.5%) Hyperlipidemia, n (%) 0.326  No 2749 (89.9%) 2450 (90.1%) 299 (88.2%)  Yes 310 (10.1%) 270 (9.9%) 40 (11.8%) Diabetes status, n (%) < 0.001  No 2887 (94.4%) 2588 (95.1%) 299 (88.2%)  Yes 131 (4.3%) 98 (3.6%) 33 (9.7%) HDL-C (mg/dL) 56.6 ± 15.9 56.9 ± 15.8 53.9 ± 17 < 0.001 LDL-C (mg/dL) 102 ± 30.2 102.1 ± 29.9 105 ± 32.3 0.320 Triglyceride (mg/dL) 85.2 ± 60.5 84.6 ± 60.2 90.8 ± 62.7 0.008 Total cholesterol (mg/dL) 177 ± 35.7 177.1 ± 35.6 180.2 ± 36.1 0.095 Lymphocyte count × 10 3 /μL 2.32 ± 0.70 2.309 ± 0.69 2.446 ± 0.73 < 0.001 Neutrophil count × 10 3 /μL 4.46 ± 1.77 4.434 ± 1.76 4.639 ± 1.82 0.034 Monocyte count × 10 3 /μL 0.544 ± 0.17 0.541 ± 0.17 0.569 ± 0.19 0.028 Platelet count × 10 3 /μL 269 ± 66.80 268.55 ± 66.87 273.77 ± 65.82 0.068 This table presents the baseline characteristics of the study population stratified by female infertility status, continuous variables are expressed as mean ± standard deviation (SD), with p-values derived from Student’s t-test. Categorical variables are reported as percentages, with p-values calculated using the weighted chi-square test OR odds ratio, CI confidence interval, SIRI Systemic Inflammatory Response Index, HOMA-IR homeostasis model assessment of insulin resistance Baseline clinical characteristics by female infertility status: insights from NHANES Data (2015–2020) This table presents the baseline characteristics of the study population stratified by female infertility status, continuous variables are expressed as mean ± standard deviation (SD), with p-values derived from Student’s t-test. Categorical variables are reported as percentages, with p-values calculated using the weighted chi-square test OR odds ratio, CI confidence interval, SIRI Systemic Inflammatory Response Index, HOMA-IR homeostasis model assessment of insulin resistance Anthropometric and metabolic characteristics showed notable disparities. Women with infertility had higher body mass index (31.72 ± 8.59 kg/m 2 vs 29.05 ± 8.05 kg/m 2 , p < 0.001), waist circumference (102.2 ± 19.85 cm vs 95.1 ± 18.50 cm, p < 0.001), and hip circumference (113 ± 18.57 cm vs 108.5 ± 16.2 cm, p < 0.001). Additionally, they exhibited elevated fasting glucose levels (107 ± 38.8 mg/dL vs 100 ± 25.2 mg/dL, p < 0.001), though HOMA-IR differences were not statistically significant (p = 0.092). While the prevalence of hypertension (p = 0.746) and hyperlipidemia (p = 0.326) did not differ significantly, diabetes was more prevalent in the infertility group (9.7% vs 3.6%, p < 0.001). Lipid profiles revealed lower HDL-C levels in the infertility group (53.9 ± 17 mg/dL) compared to the fertile group (56.9 ± 15.8 mg/dL, p < 0.001), with no significant difference observed in LDL-C or total cholesterol. Importantly, markers of systemic inflammation were significantly elevated in the infertile group. Mean lymphocyte count (2.446 ± 0.73 × 10 3 /μL, p < 0.001), neutrophil count (4.639 ± 1.82 × 10 3 /μL, p = 0.034), and monocyte count (0.569 ± 0.19 × 10 3 /μL, p = 0.028) were all significantly higher among infertile participants. Multivariate logistic regression analyses demonstrated a significant and positive association between the Systemic Inflammatory Response Index (SIRI) and female infertility (Table  2 ). In the unadjusted model (Model 1), each unit increase in SIRI was associated with a 26% increase in the odds of infertility (OR 1.26, 95% CI 1.03–1.41, p < 0.001). This association remained robust after controlling for demographic variables, including age, race, marital status, adult food security, and education level in Model 2 (OR 1.32, 95% CI 1.02–1.57, p < 0.001). In the fully adjusted model (Model 3), which incorporated behavioral and metabolic factors such as smoking, alcohol consumption, BMI, diabetes, hypertension, and hyperlipidemia, the positive relationship between SIRI and infertility persisted (OR 1.34, 95% CI 1.07–1.60, p = 0.001). Table 2 Multivariate logistic regression analysis between SIRI and female infertility Exposure OR (95% CI) (Model 1) p (Model 2) p (Model 3) p Crude model Partially adjusted model Fully adjusted model SIRI 1.26 (1.03–1.41) < 0.001 1.32 (1.02–1.57)  < 0.001 1.34 (1.07–1.60) 0.001 SIRI quartiles Q1 Ref – Ref – Ref – Q2 1.10 (0.78–1.57) 0.504 1.12 (0.80–1.64) 0.209 1.13 (0.78 -1.61) 0.517 Q3 1.22 (0.87–1.70) 0.328 1.29 (0.83–1.69) 0.411 1.37 (0.82 -1.68) 0.367 Q4 2.01 (1.36–2.44) < 0.001 2.07 (1.49–2.87) < 0.001 2.08 (1.50–2.89) < 0.001 P for trend < 0.001 < 0.001 < 0.001 The data provided includes the odds ratios (OR) and confidence intervals (95% CI) for each model, as well as the p-values indicating the statistical significance of these associations Model 1: unadjusted Model 2: adjusted for age, race, marital status, adult food security and education level Model 3: adjusted for age, race, marital status, adult food security, education level, smoking, alcohol consumption, BMI, hyperlipidemia, hypertension, diabetes status OR odds ratio, CI confidence interval, SIRI Systemic Inflammatory Response Index, BMI body mass index Multivariate logistic regression analysis between SIRI and female infertility The data provided includes the odds ratios (OR) and confidence intervals (95% CI) for each model, as well as the p-values indicating the statistical significance of these associations Model 1: unadjusted Model 2: adjusted for age, race, marital status, adult food security and education level Model 3: adjusted for age, race, marital status, adult food security, education level, smoking, alcohol consumption, BMI, hyperlipidemia, hypertension, diabetes status OR odds ratio, CI confidence interval, SIRI Systemic Inflammatory Response Index, BMI body mass index When SIRI was categorized into quartiles, a dose–response relationship was evident. Compared with women in the lowest quartile (Q1, reference), the odds of infertility rose with higher quartiles. In the unadjusted model (Model 1), the odds ratios for infertility were 1.10 (95% CI 0.78–1.57, p = 0.504) for Q2, 1.22 (95% CI 0.87–1.70, p = 0.328) for Q3, and 2.01 (95% CI 1.36–2.44, p < 0.001) for Q4. Upon adjustment for demographic covariates (Model 2), these associations became slightly stronger: Q2 (OR 1.12, 95% CI 0.80–1.64, p = 0.209), Q3 (OR 1.29, 95% CI 0.83–1.69, p = 0.411), and Q4 (OR 2.07, 95% CI 1.49–2.87, p < 0.001). In the fully adjusted model (Model 3), the odds of infertility remained significantly elevated in the highest quartile: Q2 (OR 1.13, 95% CI 0.78–1.61, p = 0.517), Q3 (OR 1.37, 95% CI 0.82–1.68, p = 0.367), and Q4 (OR 2.08, 95% CI 1.50–2.89, p < 0.001). A significant linear trend was observed across quartiles in all models (P for trend < 0.001), indicating a consistent increase in infertility risk with higher systemic inflammation. Distributional analysis across SIRI quartiles (Fig.  2 ) further substantiated the observed association between systemic inflammation and infertility status. A greater proportion of infertile women was concentrated in the highest SIRI quartile (Q4: 36%) compared to fertile women (Q4: 22%). Conversely, fertile women were more frequently found in the lowest quartile (Q2: 27%) relative to infertile women (Q1: 19%). The intermediate quartiles were comparably distributed across groups, with Q2 comprising 27% and Q3 25% among fertile women, while accounting for 22% among infertile women, respectively. These distributional disparities were statistically significant, as confirmed by the Pearson χ 2 test (χz 2  = 26.37, p = 0.001, Vcramer = 0.09), indicating a moderate divergence in SIRI-based systemic inflammatory burden between fertile and infertile women. Fig. 2 Distributional analysis by SIRI quartiles Distributional analysis by SIRI quartiles To further evaluate potential nonlinear associations between SIRI and female infertility, two flexible modeling techniques were applied: a Generalized Additive Model (GAM) and a Restricted Cubic Spline (RCS) regression (Fig.  3 ). The GAM model (Fig.  3 A) demonstrated a consistent, slightly upward curvilinear trend, indicating a progressive increase in the odds of infertility with rising SIRI values. The association became more apparent beyond SIRI values of approximately 2, with the odds ratio steadily rising from ~ 1.18 to ~ 1.32. Statistically, a significant overall association was detected (p < 0.001), though the non-linear component was not significant (p = 0.234). Similarly, the RCS analysis (Fig.  3 B) affirmed the overall positive association (p = 0.038), with the non-linearity component remaining statistically non-significant (p = 0.214). The RCS curve indicated a gradual increase in infertility odds at lower SIRI values, followed by a steeper rise beyond ~ 1,2 units, although wider confidence intervals were observed at higher SIRI levels, likely reflecting lower sample density and reduced estimation precision in that range. Fig. 3 Association between SIRI and female infertility using a GAM and RCS regression model. Graphs show ORs for female infertility according to SIRI adjusted for age, race, marital status, education level, smoking, alcohol consumption, BMI, hyperlipidemia, hypertension, and diabetes status Association between SIRI and female infertility using a GAM and RCS regression model. Graphs show ORs for female infertility according to SIRI adjusted for age, race, marital status, education level, smoking, alcohol consumption, BMI, hyperlipidemia, hypertension, and diabetes status These findings underscore a robust and dose-dependent association between systemic inflammation—as indexed by SIRI—and the risk of female infertility, observed consistently across linear and categorical models and maintained after adjustment for relevant covariates. Threshold effect analysis was conducted using a piecewise linear regression approach to investigate potential inflection points in the relationship between SIRI and female infertility (Table  3 ). The standard multivariate logistic regression model (Model a) demonstrated a statistically significant positive association between SIRI and infertility (OR 1.26, 95% CI 1.03–1.41, p < 0.001), supporting the linear trend observed in earlier analyses. Table 3 Threshold effect analysis of SIRI on female infertility using a two-piecewise linear regression model Outcomes Infertility OR (95% CIs) p value SIRI Model a (Fitting model by standard multivariate logistic regression analysis) 1.26 (1.03–1.41) < 0.001 Model b (Fitting models by two-piecewise linear regression) 1.66  Inflection point   Inflection point 1.27 (1.02, 1.44) 0.020 p for log-likelihood ratio test 0.018 The two-piecewise regression models were adjusted for age, race, marital status, education level, smoking, alcohol consumption, BMI, hyperlipidemia, hypertension, and CKD OR odds ratios, CI confidence interval, BMI body mass index, CKD chronic kidney disease Threshold effect analysis of SIRI on female infertility using a two-piecewise linear regression model The two-piecewise regression models were adjusted for age, race, marital status, education level, smoking, alcohol consumption, BMI, hyperlipidemia, hypertension, and CKD OR odds ratios, CI confidence interval, BMI body mass index, CKD chronic kidney disease However, further modeling using a two-piecewise linear regression (Model b) revealed a more complex pattern. The analysis identified an inflection point in the SIRI-infertility relationship (inflection points- exact value: 1.66), indicating a threshold-dependent effect. Below the inflection point, the association between SIRI and infertility was not statistically significant (OR 1.04, 95% CI 0.97–1.13, p = 0.112). In contrast, above the inflection point, a significantly stronger association was observed (OR 1.27, 95% CI 1.02–1.44, p = 0.020), suggesting that infertility risk increases markedly once SIRI surpasses a critical level. Importantly, the log-likelihood ratio test comparing the two models confirmed the superiority of the piecewise approach over the standard linear regression model (p for log-likelihood ratio test: 0.018), providing strong evidence for a threshold effect. These findings indicate a non-linear, inflection-based association between Systemic Inflammatory Response Index status and female infertility, whereby elevated SIRI values beyond a certain threshold confer greater reproductive risk. Subgroup analyses were conducted to examine the robustness of the association between the systemic inflammation response index (SIRI) and female infertility across various demographic and clinical strata. The findings demonstrated heterogeneity in effect estimates across subpopulations, highlighting potential effect modification by factors such as age, race/ethnicity, metabolic status, and comorbid conditions. Detailed stratified outcomes are presented in Supplementary Figure S1 . In addition, a correlation heatmap was generated to visualize the interrelationships between SIRI and key clinical and biochemical variables. The heatmap revealed distinct correlation patterns, particularly between SIRI and markers of systemic inflammation, adiposity, and metabolic dysfunction. These correlations are illustrated in Supplementary Figure S2.

SIRI was calculated using the absolute counts of neutrophils, monocytes, and lymphocytes, obtained from routine complete blood count data collected by NHANES. The following formula calculates SIRI = (Neutrophil Count × Monocyte Count)/Lymphocyte Count. In terms of Female infertility, it was characterized based on responses to the Reproductive Health Questionnaire in NHANES. Participants were asked two questions: “Have you ever sought medical advice due to difficulty in becoming pregnant?” or “Have you ever tried to get pregnant for at least a year without success?” Women who answered “yes” to at least one of them were classified as infertile. This definition is consistent with established clinical criteria for infertility diagnosis. The present study evaluated multiple covariates to control for potential confounding factors in examining the association between SIRI and female infertility. Sociodemographic and Lifestyle Factors: Demographic and lifestyle factors, including age, race/ethnicity, Marital status, educational attainment, smoking status, and alcohol consumption were categorized as we described before (2). Socioeconomic status was assessed using household income and the poverty-to-income ratio (PIR). Clinical and metabolic Factors: Body mass index (BMI, kg/m 2 ), waist and hip circumference (cm), total cholesterol (mg/dL), high-density lipoprotein cholesterol (HDL-C, mg/dL), low-density lipoprotein cholesterol (LDL-C, mg/dL), triglycerides (mg/dL), and fasting glucose (mg/dL). Hypertension was defined as a systolic blood pressure (SBP) ≥ 140 mmHg and/or diastolic blood pressure (DBP) ≥ 90 mmHg, based on measured values or self-reported diagnosis. Diabetes status was classified as either present or absent based on fasting glucose levels, HbA1C levels, self-reported history, or medication use. Hyperlipidemia was defined by triglycerides ≥ 150 mg/dL, total cholesterol ≥ 200 mg/dL, LDL-C ≥ 130 mg/dL, or HDL-C < 40 mg/dL in men and < 50 mg/dL in women, without cholesterol-lowering medications.

This Cross-Sectional study utilized data from the National Health and Nutrition Examination Survey (NHANES) for cycles conducted between 2015 and 2020. This survey employs a comprehensive protocol comprising structured interviews, standardized physical examinations, and laboratory-based assessments. All data were collected by trained personnel at specialized Mobile Examination Centers (MECs). This study investigated female participants of reproductive age (18–45 years old) who had responded to the Reproductive Health Questionnaire. A total of 3059 women were included in the final analysis after applying the following exclusion criteria: (1) those who underwent hysterectomy or bilateral oophorectomy, as defined in previous NHANES studies. It was assessed based on this question: “Have you had a hysterectomy in your past surgical history, that is, surgery to remove your uterus? And (2) those with incomplete or missing data for key variables required to calculate the Systemic Inflammation Response Index (SIRI) or determine female infertility status. Figure  1 displays a flowchart detailing the participant selection process. Fig. 1 Flow diagram of the enrollment of study participants. NHANES National Health and Nutrition Examination Survey, SIRI Systemic Inflammatory Response Index Flow diagram of the enrollment of study participants. NHANES National Health and Nutrition Examination Survey, SIRI Systemic Inflammatory Response Index All implemented procedures received ethical approval from the Research Ethics Review Board of the National Center for Health Statistics (NCHS), under Protocol #2021. Notably, written informed consent was obtained from all individuals involved in the study. The survey’s methodology, including detailed descriptions of its design and access to publicly available datasets, is thoroughly documented on the Centers for Disease Control and Prevention (CDC) NHANES website ( https://www.cdc.gov/nchs/nhanes/ ). In alignment with international research ethics, this study was conducted following the principles outlined in the Declaration of Helsinki (2013 revision) and reported following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines, ensuring methodological rigor, ethical integrity, and transparency.

This study demonstrates that the SIRI is a significant, independent predictor of female infertility in a nationally representative cohort. By integrating traditional regression analyses with advanced machine learning techniques, we identified a clear dose–response and threshold-dependent relationship between systemic inflammation and reproductive risk. These findings highlight the potential of SIRI as a valuable biomarker in infertility screening and support the incorporation of inflammatory indices into predictive models for reproductive health assessment.

This study presents a significant positive and independent association between SIRI and female infertility, even after full adjustment. In the model (Model 3), each unit increase in SIRI was associated with a 34% increase in the odds of infertility (OR 1.34, p = 0.001). It was also observed that women in the highest SIRI quartile (Q4) exhibited more than two-fold increased odds of infertility (OR 2.08, p < 0.001) compared to those in the lowest quartile, with a significant linear trend across quartiles (P for trend < 0.001). The GAM demonstrated a monotonic increase in infertility risk with rising SIRI, particularly beyond values of approximately 2.0 (Fig.  3 A), with a statistically significant overall association (p < 0.001), although the non-linearity component was not significant (p = 0.234). Similarly, the RCS regression affirmed a curvilinear pattern (p = 0.038), showing a sharper increase in infertility risk at higher SIRI levels (Fig.  3 B). These consistent results across models highlight the robustness of the SIRI-infertility association and suggest that the risk may accelerate disproportionately at moderate-to-high levels of systemic inflammation. Importantly, threshold effect analysis using piecewise linear regression identified a critical inflection point at SIRI = 1.66. Below this threshold, the association was not statistically significant (OR 1.04, p = 0.112), whereas above it, the odds of infertility increased significantly (OR 1.27, p = 0.020). The superiority of the piecewise model was confirmed via a log-likelihood ratio test (p = 0.018), indicating a threshold-dependent relationship. These findings suggest that a minimal inflammatory burden may be tolerated without impairing fertility, but surpassing this threshold could initiate pathophysiological processes that disrupt ovulation, endometrial receptivity, or immune tolerance required for conception. The biological plausibility of these findings is supported by the significant elevation in inflammatory blood cell counts (lymphocytes, neutrophils, monocytes) among infertile women, as observed in the descriptive analyses (p-values < 0.05), although the mean SIRI itself did not differ significantly between groups at baseline (p = 0.219). This underscores the value of modeling SIRI as a continuous variable and exploring its effects through advanced statistical techniques. Finally, machine learning validation confirmed SIRI as a top-tier predictor of infertility. Through ensemble learning techniques—especially XGBoost, which achieved the highest AUC (0.866)—SIRI consistently ranked among the most influential features, alongside age, BMI, diabetes, hypertension, and total cholesterol. The model’s robust performance across training and validation sets, reinforced by SHAP-based interpretability, revealed that higher SIRI values substantially contributed to predicted infertility risk. These insights support the clinical relevance of incorporating SIRI into future risk stratification models for female reproductive health. SIRI is a promising inflammatory marker integrating neutrophil, lymphocyte, and monocyte counts. Hence, it can potentially reflect two distinct biological pathways: the inflammatory response and the adaptive immune response. To our knowledge, emerging evidence has indirectly addressed the association between SIRI and female infertility; hence, the correlation between SIRI and infertility is still unclear. Inflammation and immunity are widely believed to have a well-established role in various causes of female infertility, such as PCOS and PID, and even unexplained infertility, thereby new emerging indicators derived from CBC, like SIRI, SII, and NLR, have been linked to various pregnancy complications and miscarriage (2). Neutrophil counts are one of the components of the SIRI formula [ 14 ], The authors propose that elevated neutrophil levels may play a role in contributing to infertility among women diagnosed with PCOS, which is consistent with findings documented by Orio et al. [ 7 , 15 ]. This claim is further supported by Hu et al., who found that PCOS patients experienced high levels of inflammation, evidenced by higher concentrations of tumor necrosis factor (TNF)-α and monocyte chemoattractant protein (MCP)-1 [ 16 , 17 ]. Lymphocytes represent another key component of the SIRI. Evidence suggests that increased lymphocyte counts, along with the enhanced secretion of proinflammatory cytokines such as tumor necrosis factor-α (TNF-α) and interleukin-6 (IL-6), may exacerbate inflammatory processes in patients with PCOS [ 18 , 19 ]. In a retrospective case–control study by Sahin et al., pregnant woman with systemic lupus erythematosus (SLE) exhibited significantly higher levels of inflammatory biomarkers, including SIRI, compared with pregnant women without SLE [ 20 ]. Moreover, when pregnant women with SLE were stratified according to the presence of perinatal complications, those who experienced perinatal complications demonstrated significantly elevated SIRI levels [ 21 ]. Furthermore, a retrospective cohort study found that elevated SIRI levels were associated with an increased risk of subsequent preeclampsia, in line with the findings reported by Seyhanli et al. [ 22 ]. Moreover, elevated inflammatory markers were found to negatively impact not only the total number of oocytes, but also their quality through proinflammatory cytokines and reactive oxygen species [ 23 – 26 ]. To address this issue, several studies observed that treatment with antioxidants could alleviate inflammation and might improve fertility outcomes in infertile women with high inflammatory burden [ 27 , 28 ]. Given the literature, SIRI is not only a precise predictive inflammatory marker, but also a reliable, non-invasive, easy-to-access, and cost-effective method, which is universally available by obtaining a simple CBC test. SIRI provides physicians with insight into a patient’s inflammatory and immune response. Previously, it was assumed that it might be a helpful paraclinical test for oncologists or infectious disease specialists [ 29 ]. Nevertheless, recent developments have altered the situation. Lin et al. reported that higher SIRI levels were associated with the presence of atrial fibrillation (AF) in patients with ischemic stroke, were also linked to increased financial burden and poorer clinical outcome [ 30 ]. Interestingly, finding from a prospective cohort study indicated that elevated SIRI are associated with a higher risk of incident hypertension [ 31 ]. Consistently, Hui et al. observed that among individuals with HIV, higher SIRI levels were linked to an increased likelihood of developing hypertension [ 32 ]. Additionally, Wang et al. identified a significant association between SIRI and metabolic disease, and higher SIRI levels were associated with an increased risk of cardiovascular disease (CVD) [ 33 ]. Moreover, a study by Xia et al. demonstrated that novel inflammatory markers like SIRI are closely linked with cardiovascular and all-cause mortality [ 34 ]. In line with these observations, a nationwide cross-sectional study identified SIRI as an independent risk factor for CVD prevalence, particularly among individuals with obesity [ 35 ]. Another study also underscored the potential role of SIRI in predicting disease severity and acute kidney injury (AKI) among patients with pancreatitis [ 36 ]. In a study by Okutan et al., SIRI was shown to have potential as a complementary diagnostic marker alongside established markers such as rheumatoid factor (RF) and anti-CCP antibodies. Moreover, SIRI levels were positively correlated with disease severity [ 37 ]. In conclusion, based on our results, another clinical significance of SIRI is as a screening tool for female infertility; individuals with the highest SIRI quartile (Q4) had a higher infertility rate, and it can probably be a more accurate predictor for patients with PCOS. These findings contribute to the identification of female infertility risk factors, which may inform the development of future preventive strategies to mitigate the adverse effects of infertility and its social burden. As we stated earlier, there is a limited number of studies that have indirectly addressed the association between SIRI and female infertility; hence, the correlation between SIRI and infertility is still unclear. Our study possesses the merit to directly explore this through both traditional regression models and advanced machine learning techniques reinforces the credibility. Moreover, interpretable methods such as XGBoost and SHAP analysis were leveraged to refine the assessment of SIRI’s predictive performance and the dynamic interplay among influential variables, yielding insights with both analytical depth and clinical value. However, the current study had several limitations. First, owing to the cross-sectional design, we could not establish definitive causal relationships between SIRI and infertility, leaving it uncertain whether elevated SIRI levels contribute to the development of infertility or merely reflect a secondary response to underlying reproductive dysfunction that influences systemic inflammatory processes. Second, failure to adjust for potential confounders such as dietary habits, psychosocial stress, or environmental exposures may have affected the observed association between SIRI and female infertility. Third, infertility status was obtained by questionnaire surveys, rather than clinically confirmed diagnoses, which may have introduced recall bias. Last but not least is the absence of direct identification of specific reproductive conditions, such as PCOS or endometriosis, that could have optimized our models.

The statistical methodology adhered to the analytical guidelines established by the NCHS and CDC. All analyses incorporated the appropriate sampling weights, stratification variables, and primary sampling units provided by NHANES to ensure nationally representative estimates. Continuous variables were summarized as weighted means with standard deviations, while categorical variables were presented as weighted frequencies and percentages. Group comparisons were performed using chi-square tests for categorical variables and independent-sample t-tests or the Mann–Whitney U test or Wilcoxon rank-sum test for continuous variables, as appropriate. The association between SIRI and female infertility was evaluated using weighted multivariable logistic regression models. SIRI was assessed as both a continuous and a categorical variable based on quartile classification. The median value of each quartile was modeled as a continuous variable to assess for linear trends. Results are reported as odds ratios (ORs) with corresponding 95% confidence intervals (CIs). To capture potential nonlinear associations, generalized additive models (GAMs) with penalized splines and restricted cubic spline (RCS) regression were employed. Three regression models with increasing levels of adjustment were developed to systematically control for confounders, as we had done before (2). All statistical analyses were conducted using R software (version 4.2.1) and EmpowerStats (version 4). A two-sided p-value of < 0.05 was considered statistically significant. Complementary machine learning analyses were carried out using Python (version 3.13.3), with model development and validation performed using established libraries including scikit-learn, TensorFlow, and Keras.

Despite remarkable achievements and cutting-edge science in medicine, female infertility has still been a serious health concern over the past couple of decades, highlighted by a significant rise in both prevalence and related disease burden. Beyond its medical implications, it frequently poses serious psychosocial consequences that affect all aspects of life [ 1 ]. Female infertility is identified as the failure to conceive in a childbearing age woman following 12 months or more of regular, unprotected sexual intercourse. Clinically, it is classified as either primary, which is characterized by the absence of any prior conception, or secondary, which indicates an inability to conceive again after at least one previous pregnancy [ 2 , 3 ]. Female-implicated infertility causes are notably accounting for approximately 35–50% of all cases. The etiology of female infertility is multifactorial, involving conditions that may impair ovulation, hormonal dysregulation, fallopian tube and uterine structural abnormalities. Moreover, other possibilities are endometriosis, polycystic ovarian syndrome (PCOS), premature ovarian failure, and pelvic inflammatory disease (PID) [ 4 , 5 ]. Some etiologies, such as PCOS, endometriosis, and PID, underscore the key role of inflammation and immune-endocrine dysregulation in the pathogenesis of female infertility [ 6 – 8 ]. Recently, there has been rising interest in exploring inflammatory markers based on complete blood count (CBC) data. CBC and parameters derived from it, neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and lymphocyte-to-monocyte ratio (LMR), can be a valuable indicator of inflammation and potential in predicting disease status and clinical prognosis [ 9 , 10 ]. Systemic Inflammation Response Index (SIRI) is a blood-based biomarker widely used to predict comprehensive clinical outcomes in various medical conditions. It integrates neutrophil, monocyte, and lymphocyte counts to provide a more comprehensive assessment of overall inflammatory status compared to traditional single-cell biomarkers. Additionally, it has shown strong predictive value as a prognostic biomarker in various diseases [ 11 – 13 ]. The present study aims to assess the predictive role of SIRI for female infertility by applying a hybrid approach that compares regression models and machine learning algorithms using nationally representative NHANES data.

Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 501 KB) Supplementary file1 (DOCX 501 KB)