Development and temporal validation of a gradient boosting machine model for endometriosis using routine laboratory indicators

In: Frontiers in Medicine · 2026 · vol. 13 · doi:10.3389/fmed.2026.1934454 · W7236086355
OA: gold CC0
⚙ AI-generated summary by qwen3.7-flash, 2026-11-20 ⓘ

A gradient boosting machine model using eight routine laboratory indicators demonstrated consistent discrimination for endometriosis identification but showed limited sensitivity and no significant improvement over extended clinical models.

One-sentence paraphrase of the abstract; not a substitute for reading it. No clinical advice. How this works

⚙ AI-generated deep summary by qwen3.7-flash, 2026-11-20 · read from full text ⓘ

This retrospective study developed and validated a gradient boosting machine model to identify endometriosis using eight routine laboratory indicators, including lipid profiles, complete blood counts, and CA125. The model demonstrated high discrimination with temporal validation area under the curve scores of 0.8565, though it showed limited sensitivity and no statistically significant improvement over extended clinical models. The authors note that while the algorithm is reproducible, further multicenter external validation is required before clinical implementation can be considered. This paper is centrally about endometriosis — specifically focusing on developing a non-invasive diagnostic tool using routine blood tests to aid in identifying the disease among women undergoing gynecological evaluation.

Read from the paper's body, not the abstract. Not a substitute for reading the paper. No clinical advice. How this works

Abstract

Background Endometriosis (EM) is a chronic gynecological disorder for which timely diagnosis remains challenging. Routine laboratory indicators may inform EM identification, but their performance in multivariable diagnostic models requires rigorous validation. This study aimed to develop and temporally validate a machine learning model for identifying EM among women undergoing gynecological evaluation. Methods This retrospective, single-center study included 2,452 women (839 with EM and 1,613 disease controls). Participants were allocated to training ( n = 1,029), internal test ( n = 441), and temporal validation ( n = 982) cohorts. Twelve candidate predictors were evaluated using least absolute shrinkage and selection operator regression and Boruta. Six algorithms were compared using repeated nested cross-validation. The selected model underwent discrimination, calibration, exploratory decision curve, clinical comparator, and sensitivity analyses. Results Eight indicators were selected: triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, lymphocyte count, neutrophil count, monocyte count, platelet count, and CA125. Gradient boosting machine (GBM) achieved the highest mean area under the receiver operating characteristic curve (AUC) during repeated nested cross-validation (0.8483). The GBM achieved AUCs of 0.8401 (95% CI, 0.8006–0.8797) internally and 0.8565 (95% CI, 0.8320–0.8810) temporally. At a fixed threshold of 0.50, sensitivity was 55.63% and 59.52%, while specificity was 91.38% and 88.85%, respectively. AUC differences vs. the extended clinical model were not statistically significant (DeLong P = 0.758 and P = 0.294, respectively). The publicly released frozen GBM pipeline reproduced stored predictions to numerical precision in 987 participants with complete predictor measurements. Conclusion An eight-indicator GBM demonstrated consistent discrimination across internal and temporal validation cohorts. However, improvement over the extended clinical model was not statistically significant, and sensitivity was limited. Routine laboratory-based models warrant further investigation as potential adjuncts to gynecological assessment. Multicenter external validation and prospective clinical utility evaluation are required before implementation.
Full text 76,298 characters · extracted from oa-doi-fallback · 6 sections · click to expand

Abstract

Background: Endometriosis (EM) is a chronic gynecological disorder for which timely diagnosis remains challenging. Routine laboratory indicators may inform EM identification, but their performance in multivariable diagnostic models requires rigorous validation. This study aimed to develop and temporally validate a machine learning model for identifying EM among women undergoing gynecological evaluation.

Methods

This retrospective, single-center study included 2,452 women (839 with EM and 1,613 disease controls). Participants were allocated to training (n = 1,029), internal test (n = 441), and temporal validation (n = 982) cohorts. Twelve candidate predictors were evaluated using least absolute shrinkage and selection operator regression and Boruta. Six algorithms were compared using repeated nested cross-validation. The selected model underwent discrimination, calibration, exploratory decision curve, clinical comparator, and sensitivity analyses.

Results

Eight indicators were selected: triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, lymphocyte count, neutrophil count, monocyte count, platelet count, and CA125. Gradient boosting machine (GBM) achieved the highest mean area under the receiver operating characteristic curve (AUC) during repeated nested cross-validation (0.8483). The GBM achieved AUCs of 0.8401 (95% CI, 0.8006–0.8797) internally and 0.8565 (95% CI, 0.8320–0.8810) temporally. At a fixed threshold of 0.50, sensitivity was 55.63% and 59.52%, while specificity was 91.38% and 88.85%, respectively. AUC differences vs. the extended clinical model were not statistically significant (DeLong P = 0.758 and P = 0.294, respectively). The publicly released frozen GBM pipeline reproduced stored predictions to numerical precision in 987 participants with complete predictor measurements.

Conclusion

An eight-indicator GBM demonstrated consistent discrimination across internal and temporal validation cohorts. However, improvement over the extended clinical model was not statistically significant, and sensitivity was limited. Routine laboratory-based models warrant further investigation as potential adjuncts to gynecological assessment. Multicenter external validation and prospective clinical utility evaluation are required before implementation. 1 Introduction Endometriosis (EM) is a common chronic gynecological disorder characterized by the presence of endometrial-like tissue outside the uterus, affecting approximately 10% of women of reproductive age ( , ). Its clinical manifestations are heterogeneous and commonly include dysmenorrhea, chronic pelvic pain, dyspareunia, and infertility ( , ). EM can substantially impair quality of life and impose considerable healthcare and productivity-related costs ( ). Despite its clinical burden, diagnosis is frequently delayed because symptoms are often nonspecific and overlap with those of other gynecological disorders ( , ). A systematic scoping review further highlighted substantial variation in diagnostic delays across healthcare systems and patient populations ( ).Current diagnostic evaluation for suspected EM integrates clinical history, physical examination, and imaging, particularly ultrasonography and magnetic resonance imaging ( ). Contemporary guidelines do not require laparoscopy and histological confirmation as mandatory first-line diagnostic procedures for all patients. Imaging is recommended as part of the diagnostic work-up, whereas laparoscopy may be considered when imaging is negative and empirical treatment is unsuccessful or inappropriate ( ). Nevertheless, diagnostic performance varies according to lesion phenotype, anatomical location, and imaging expertise. Negative imaging does not exclude EM, particularly superficial peritoneal disease ( , ). These limitations highlight the need for accessible diagnostic approaches that complement established clinical assessment.Blood-based biomarkers have attracted considerable interest because of their accessibility and potential to reflect multiple biological processes associated with EM. Cancer antigen 125 (CA125) is among the most extensively investigated circulating biomarkers for EM. Although CA125 has demonstrated relatively high specificity, its sensitivity is limited, particularly in patients with minimal disease ( ). Furthermore, available evidence does not support the use of circulating biomarkers as standalone diagnostic tests for EM ( , ). A systematic review of candidate biomarkers across biological compartments also highlighted substantial heterogeneity in biomarker performance and the absence of a universally validated noninvasive diagnostic marker ( ). Nevertheless, CA125 remains an important component of biomarker-based research and may provide complementary diagnostic information when integrated with other routinely available clinical indicators.Beyond CA125, systemic inflammatory and metabolic alterations may provide additional information relevant to EM identification. EM involves complex interactions among innate and adaptive immune responses, including neutrophils, macrophages, natural killer cells, and lymphocyte populations ( ). Neutrophils have been implicated in inflammatory signaling and angiogenesis within the systemic and pelvic microenvironments associated with EM ( ). Accordingly, routinely measured hematological parameters may capture aspects of systemic inflammatory responses, although these indicators are not specific to EM. Alterations in circulating lipid profiles have also been reported among women with EM, although findings for individual lipid parameters remain inconsistent across populations ( ). Lipidomic and metabolomic investigations have further identified disease-associated metabolic signatures ( , ). However, the predictive contribution of conventional hematological and lipid measurements requires evaluation within appropriately developed and validated multivariable models.Machine learning provides a framework for integrating multiple clinical and laboratory variables and modeling potentially nonlinear relationships among predictors. Previous studies have explored machine learning approaches for EM diagnosis using combinations of clinical and laboratory indicators ( ). However, diagnostic biomarker models may show reduced performance during independent validation, underscoring the importance of rigorous evaluation beyond model development ( ). A recent systematic review and meta-analysis reported promising diagnostic performance of machine learning approaches for EM while identifying substantial methodological heterogeneity ( ). Moreover, favorable discrimination alone does not establish clinical usefulness. Prediction models should also be evaluated for calibration, performance relative to clinically relevant comparators, and potential consequences of their use in decision-making ( – ).Therefore, this retrospective study aimed to develop and temporally validate a machine learning–based diagnostic model for identifying existing EM among women undergoing gynecological evaluation using routinely available laboratory indicators. Candidate predictors were evaluated through feature-selection procedures, and six machine learning algorithms were compared using repeated nested cross-validation within the training dataset. The selected model was subsequently assessed in a held-out internal test set and an independent temporal validation cohort. In addition to evaluating discrimination and calibration, we compared its performance with CA125 alone and clinical reference models incorporating demographic characteristics, reproductive history, symptoms, and ultrasonographic findings. Exploratory decision curve analysis was performed to assess cohort-specific net benefit across a range of decision thresholds. The objective was to investigate whether integrating routine laboratory indicators could provide complementary diagnostic information within existing gynecological assessment rather than replace established clinical evaluation. 2 Methods 2.1 Study design and participants This retrospective diagnostic modeling study was conducted at the First Affiliated Hospital of Bengbu Medical University. Medical records of women undergoing gynecological evaluation between January 2020 and December 2025 were retrospectively reviewed. A total of 23,606 patient records were initially assessed for eligibility. Patients were sequentially excluded according to the following criteria: (1) age younger than 18 years or older than 55 years (n = 8,658); (2) a final gynecological diagnosis outside the predefined study population (n = 3,924); (3) concomitant malignant disease (n = 3,636); (4) active systemic autoimmune disease (n = 683); (5) use of lipid-lowering medications within 90 days before the initial blood sampling after hospital admission (n = 1,685); and (6) antibiotic or immunomodulatory medication use before the initial blood sampling after hospital admission (n = 2,568). These exclusion criteria were applied sequentially, with each patient counted only once. The final study population comprised 2,452 eligible women (Figure 1). Figure 1 Patients evaluated between January 2020 and December 2023 constituted the development-period cohort (n = 1,470; 503 EM cases and 967 disease controls). This cohort was stratified by EM status and randomly divided into a training set (n = 1,029; 352 EM cases and 677 disease controls) and an internal test set (n = 441; 151 EM cases and 290 disease controls). An independent temporal validation cohort comprised 982 patients evaluated between January 2024 and December 2025 (336 EM cases and 646 disease controls). The internal test set and temporal validation cohort were reserved for independent model evaluation and were not used for predictor selection, hyperparameter tuning, or algorithm selection. No fixed case-to-control ratio was imposed, and no individual case-control matching was performed. The similar ratios across study periods arose from the eligible populations after sequential exclusions. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the First Affiliated Hospital of Bengbu Medical University (Approval No. [2023]449). The requirement for individual informed consent was waived by the Ethics Committee because of the retrospective nature of the study. 2.2 Outcome definition and disease-control selection The primary diagnostic outcome was the presence or absence of existing endometriosis (EM). EM cases were identified based on surgical and histopathological confirmation. The disease-control group comprised women hospitalized during the corresponding study periods for other gynecological conditions, including adenomyosis, ovarian cysts, pelvic inflammatory disease, uterine fibroids, functional dysmenorrhea, chronic pelvic pain, and unexplained infertility. Disease-control diagnoses were established using available clinical, operative, and pathological records. For patients who did not undergo surgery, the final gynecological diagnosis was determined from presenting symptoms and ancillary examinations. Most disease-control participants underwent surgery for their primary gynecological conditions. Patients with documented concomitant EM were not classified as disease controls. Because surgical verification was not uniformly available for all disease-control participants, occult EM could not be definitively excluded in every case. Accordingly, the comparator group was designated as a disease-control population rather than a uniformly surgically verified non-EM population. Across the training, internal test, and temporal validation partitions, 522/677 (77.1%), 219/290 (75.5%), and 505/646 (78.2%) disease controls, respectively, had a recorded surgical evaluation. Surgical evaluation was not assumed to exclude all occult EM. 2.3 Candidate predictors and data collection Twelve candidate predictors were prespecified based on their routine clinical availability and potential relevance to EM. These included age, body mass index (BMI), total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), lymphocyte count (LYM), neutrophil count (NEUT), monocyte count (MONO), white blood cell count (WBC), platelet count (PLT), and cancer antigen 125 (CA125). Age and BMI were obtained from admission records. Laboratory variables were extracted from the first blood sample collected after hospital admission and before definitive treatment or surgical intervention. To minimize the potential influence of treatment-related changes on circulating biomarkers, patients who had received lipid-lowering medications within 90 days before blood sampling or antibiotic or immunomodulatory medications before the initial blood sampling were excluded according to the predefined eligibility criteria. All included patients underwent fasting blood sampling. CA125 was measured using an electrochemiluminescence immunoassay on the Roche cobas e 801 analyzer (Roche Diagnostics) and reported in U/mL. The institutional CA125 testing platform remained unchanged during the study period. Blood sampling was not standardized to a specific follicular or luteal phase; however, samples were collected outside the menstrual period to reduce potential menstruation-related fluctuations in circulating biomarkers, particularly CA125. 2.4 Data preprocessing Missing predictor values were handled using multiple imputation by chained equations (MICE), with five imputed datasets generated for feature selection and model development. Predictive mean matching was used for continuous predictors. Within the repeated nested cross-validation framework, imputation was incorporated into the resampling procedure, with the corresponding training partitions used to estimate imputation models and held-out assessment observations excluded from imputation-model fitting. For independent evaluation, the original MICE workflow used training observations to estimate the imputation models, while observations from the internal test set or temporal validation cohort were excluded from fitting the imputation models. Held-out predictor measurements could be supplied for conditional imputation, but held-out outcomes were not used to estimate imputation models. The procedure was implemented within the original analysis pipeline rather than as a separately fitted, portable imputation object. The independent evaluation cohorts were not used for predictor selection, algorithm selection, or hyperparameter tuning. Continuous predictors were winsorized at the 1st and 99th percentiles to limit the influence of extreme observations. Standardization was applied where required by the modeling algorithm. Winsorization limits and standardization parameters were estimated from the relevant training partitions and subsequently applied to their corresponding assessment data. The final GBM models retained their training-derived preprocessing parameters for independent evaluation and subsequent reproducible prediction. The imputed datasets were used for predictive modeling rather than formal multiple-imputation inference; consequently, Rubin's rules were not applied to pool model coefficients or inferential estimates. No synthetic oversampling or other class-balancing procedure was used. 2.5 Feature selection and stability assessment Twelve candidate predictors were initially considered for feature selection using least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. LASSO selection used the one-standard-error penalty criterion (λ.1se), while Boruta-confirmed features were used to identify predictors supported by the random forest–based selection procedure. To reduce redundancy between highly correlated predictors, WBC was excluded from the eligible feature set when its absolute Pearson correlation with NEUT exceeded 0.90 in the corresponding training data. The intersection of LASSO-selected and Boruta-confirmed predictors was used to determine joint selection status. Feature-selection stability was first examined descriptively across 50 outer resampling iterations, corresponding to five repeats of 10-fold cross-validation within the training dataset. Selection frequencies were summarized separately for LASSO, Boruta, and their intersection (Figure 2A; Supplementary Table S1). A descriptive frequency reference threshold of 70% was used to characterize selection stability. Four predictors—CA125, HDL, LDL, and NEUT—reached this threshold for the intersection of LASSO and Boruta. A threshold sensitivity summary at 50%, 60%, 70%, and 80% is provided in Supplementary Table S6; the same four intersection predictors were retained at each threshold. This descriptive stability assessment did not determine the final eight-predictor model. Figure 2 Final predictor selection was conducted separately using five multiply imputed versions of the complete training dataset. LASSO (λ.1se) and Boruta were applied to each imputed dataset, and predictors selected by both methods in at least three of the five datasets were retained through majority voting. This procedure identified eight predictors: TG, HDL, LDL, LYM, NEUT, MONO, PLT, and CA125 (Figure 2B). 2.6 Model development and algorithm selection Six candidate machine learning algorithms were compared: logistic regression (LR), support vector machine (SVM), gradient boosting machine (GBM), neural network (NN), extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost). Algorithm comparison was performed exclusively within the training dataset (n = 1,029) using five repeats of nested 10-fold cross-validation, with five-fold cross-validation for hyperparameter tuning in each outer training partition. Imputation, preprocessing, feature selection, and hyperparameter optimization were conducted within the relevant training folds. Predictors selected in an outer training partition were not reused as a fixed feature set during inner-fold tuning; feature selection was repeated within the inner training folds to preserve separation between tuning and assessment. Discrimination was summarized using the mean area under the receiver operating characteristic curve (AUC) across the five outer cross-validation repeats. The algorithm with the highest mean development-stage AUC was selected for final model development. GBM achieved the highest mean AUC (0.8483), followed closely by XGBoost (0.8470), AdaBoost (0.8397), neural network (0.8351), SVM (0.8275), and logistic regression (0.8113) (Figure 3). The small difference between GBM and XGBoost was considered descriptive and was not interpreted as evidence of statistically significant superiority. Figure 3 Following development-stage algorithm selection, the final GBM configuration was determined using training-data-only tuning and subsequently fitted across five multiply imputed training datasets. The final configuration comprised 150 trees, an interaction depth of three, a shrinkage rate of 0.05, a minimum terminal-node size of 10, and a bag fraction of 0.8. The predicted probabilities from the five fitted GBM models were averaged to obtain the final prediction score. For computational reproducibility, the five fitted GBM models and their frozen training-derived preprocessing parameters were preserved without refitting. The publicly released standalone prediction function requires complete numeric measurements of the eight selected predictors and does not perform missing-data imputation. Therefore, its applicability to patients with incomplete predictor measurements has not been established. 2.7 Model performance evaluation The final GBM model was evaluated in the held-out internal test set (n = 441) and the independent temporal validation cohort (n = 982). Neither dataset was used for predictor selection, algorithm selection, or hyperparameter optimization. Model discrimination was assessed using receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC), with 95% confidence intervals (CIs). Classification performance was assessed using sensitivity and specificity at the fixed classification threshold of 0.50. To assess performance relative to clinically relevant benchmarks, the GBM was compared with three reference models: CA125 alone; a basic clinical model incorporating CA125, age, BMI, and previous pregnancy history; and an extended clinical model additionally incorporating dysmenorrhea, chronic pelvic pain, infertility, and transvaginal ultrasonography (TVUS) findings suggestive of EM. Symptoms and TVUS findings were documented before the final diagnostic determination. Missing values in the clinical reference models were handled using training-derived median imputation for continuous variables and mode imputation for categorical variables. This preprocessing approach differed from the MICE-based strategy used for the GBM model. All clinical reference models were evaluated in the same internal test and temporal validation datasets (Figure 4; Table 2). Figure 4 Differences in AUC between the GBM and the reference models were assessed using paired DeLong tests. Particular emphasis was placed on the comparison with the extended clinical model, which incorporated a broader range of routinely available clinical information. Model comparisons were interpreted according to both the magnitude of AUC differences and their statistical uncertainty. 2.8 Classification threshold A fixed predicted-probability threshold of 0.50 was used to classify patients as having or not having EM. This threshold was applied consistently in the internal test set and temporal validation cohort without re-optimization in either dataset. Sensitivity and specificity were calculated at this threshold for the primary independent validation analyses. Positive predictive value (PPV) and negative predictive value (NPV) were additionally evaluated in the complete-predictor validation subsets. Because the observed proportion of EM cases in the study cohorts may not represent disease prevalence in the intended clinical population, PPV and NPV were interpreted as cohort-specific performance measures rather than directly generalizable population estimates. 2.9 Calibration and clinical utility Calibration of the final GBM model was assessed separately in the internal test set and temporal validation cohort. Calibration plots were constructed by comparing observed EM proportions with mean predicted probabilities across ten quantile-based groups. Numerical calibration measures included the calibration intercept, calibration slope, and Brier score. The calibration intercept and slope were used to evaluate systematic prediction bias and agreement between predicted and observed risk gradients, respectively, while the Brier score quantified overall prediction error. Potential clinical utility was explored using decision curve analysis (DCA) in each independent evaluation cohort (Figure 5). The GBM model was compared with treat-all and treat-none strategies across a range of threshold probabilities using the original, unrecalibrated GBM predictions. Net benefit was calculated using the observed EM case proportions in the corresponding study cohorts. Figure 5 Because the case proportions in this retrospective study may not reflect the prevalence of EM in the intended clinical population, DCA was interpreted as an exploratory assessment of cohort-specific net benefit rather than definitive evidence of population-level clinical utility. No recalibration to the baseline disease probability of a defined target population was performed for the primary decision curves. 2.10 Sensitivity and subgroup analyses Age-adjusted exploratory associations between EM and TG, HDL, or LDL were assessed using separate logistic regression models with age as a covariate, restricted to observations with complete lipid and age measurements. Odds ratios (ORs) were expressed per one cohort-specific standard deviation increase in each lipid measure. These association analyses were not used for GBM selection or refitting. Several additional analyses were conducted to assess the robustness of model performance. A sensitivity analysis using median imputation was performed by refitting the GBM model under an alternative missing-data handling strategy and evaluating its discrimination in the internal test and temporal validation cohorts. Subgroup analyses examined model discrimination in surgically evaluated disease controls and across clinically relevant EM subgroups, including revised American Society for Reproductive Medicine (rASRM) disease stage and the presence or absence of ovarian endometrioma. Additional exploratory analyses were performed according to available clinical characteristics, including TVUS findings where applicable. Subgroup AUCs were interpreted descriptively. Because some EM subgroups were compared against overlapping disease-control populations, subgroup sample sizes were not considered mutually exclusive. No claims of statistically significant between-subgroup differences were made without formal interaction testing. To evaluate the performance of the publicly released prediction tool in its intended complete-predictor population, an additional complete-predictor analysis was conducted among participants with observed measurements of all eight selected predictors. The original fitted GBM models and frozen preprocessing parameters were applied without refitting or recalibration. Discrimination was assessed using AUC with 95% confidence intervals. Calibration was evaluated using the Brier score, calibration intercept, and calibration slope. Sensitivity, specificity, PPV, and NPV were calculated at the fixed classification threshold of 0.50. Predictions generated by the standalone function were compared with the corresponding stored model predictions for participants with complete measurements of all eight selected predictors. Numerical reproducibility was assessed using the maximum absolute difference between paired predicted probabilities. The results of the complete-predictor performance and reproducibility analyses are reported in Supplementary Table S5. 2.11 Statistical analysis Continuous variables were summarized as mean ± standard deviation (SD) or median (interquartile range [IQR]), as appropriate. Categorical variables were presented as counts and percentages. Between-group comparisons were performed using suitable parametric or nonparametric tests for continuous variables and chi-square or Fisher's exact tests for categorical variables, as appropriate. Standardized mean differences (SMDs) were used where applicable to describe differences in baseline characteristics between study cohorts. Univariable comparisons were descriptive and were not used as criteria for predictor selection. AUCs were estimated from independent evaluation predictions, and paired DeLong tests were used for comparisons between models evaluated in the same participants. All statistical tests were two-sided, with P < 0.05 considered statistically significant. Repeated nested cross-validation results were summarized descriptively; variability across cross-validation repeats was not interpreted as a patient-independent confidence interval. This retrospective study included all eligible participants during the predefined study periods, and no prospective sample-size calculation was performed. Statistical analyses and model development were conducted in R. 2.12 Reporting guideline The study was reported with reference to the TRIPOD + AI framework for clinical prediction models ( ). The reporting checklist was used to guide transparent description of the study population, predictor handling, model development, validation, performance assessment, and limitations.3 Results 3.1 Study population and baseline characteristics A total of 23,606 patient records were initially screened between January 2020 and December 2025. After sequential application of the predefined exclusion criteria, 2,452 eligible women were included in the final analysis (Figure 1). Among these participants, 839 were diagnosed with endometriosis (EM), and 1,613 were classified as disease controls. The development-period cohort (January 2020 to December 2023) comprised 1,470 women, including 503 patients with EM and 967 disease controls. Following stratified random allocation, 1,029 participants (352 EM cases and 677 disease controls) were assigned to the training set, and 441 participants (151 EM cases and 290 disease controls) were assigned to the held-out internal test set. The independent temporal validation cohort (January 2024 to December 2025) included 982 women, comprising 336 EM cases and 646 disease controls. The proportion of EM cases was approximately 34.2% in each cohort. Baseline demographic characteristics and routinely measured laboratory indicators, including hematological parameters, lipid profiles, and CA125 levels, are summarized in Table 1. The internal test set and temporal validation cohort were not used during model development or algorithm selection. Table 1 | Characteristic | Statistic/category | Training (n = 1,029) | Internal test (n = 441) | Temporal validation (n = 982) | |---|---|---|---|---| | Age (years) | Median [Q1, Q3] | 34.00 [28.00, 40.00] | 33.00 [27.00, 40.00] | 34.00 [28.00, 40.00] | | Missing, n (%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | | | Body mass index (kg/m2) | Median [Q1, Q3] | 22.20 [20.30, 24.10] | 22.20 [20.30, 24.30] | 21.90 [20.10, 23.80] | | Missing, n (%) | 44 (4.3%) | 17 (3.9%) | 49 (5.0%) | | | CA125 (U/mL) | Median [Q1, Q3] | 21.04 [13.18, 36.75] | 20.10 [13.03, 37.80] | 21.10 [13.30, 37.35] | | Missing, n (%) | 62 (6.0%) | 29 (6.6%) | 95 (9.7%) | | | Total cholesterol (mmol/L) | Median [Q1, Q3] | 4.38 [3.88, 5.06] | 4.42 [3.92, 5.09] | 4.44 [3.92, 5.00] | | Missing, n (%) | 51 (5.0%) | 27 (6.1%) | 69 (7.0%) | | | Triglycerides (mmol/L) | Median [Q1, Q3] | 1.18 [0.92, 1.56] | 1.16 [0.92, 1.53] | 1.19 [0.93, 1.56] | | Missing, n (%) | 51 (5.0%) | 27 (6.1%) | 69 (7.0%) | | | HDL cholesterol (mmol/L) | Median [Q1, Q3] | 1.29 [1.11, 1.49] | 1.31 [1.15, 1.51] | 1.29 [1.11, 1.49] | | Missing, n (%) | 51 (5.0%) | 27 (6.1%) | 69 (7.0%) | | | LDL cholesterol (mmol/L) | Median [Q1, Q3] | 2.54 [2.12, 3.04] | 2.59 [2.13, 3.14] | 2.57 [2.18, 3.12] | | Missing, n (%) | 51 (5.0%) | 27 (6.1%) | 69 (7.0%) | | | White blood cells (×109/L) | Median [Q1, Q3] | 6.21 [5.33, 7.05] | 6.25 [5.30, 7.31] | 6.22 [5.45, 7.16] | | Missing, n (%) | 42 (4.1%) | 22 (5.0%) | 53 (5.4%) | | | Neutrophils (×109/L) | Median [Q1, Q3] | 3.46 [2.83, 4.17] | 3.53 [2.87, 4.33] | 3.51 [2.92, 4.26] | | Missing, n (%) | 40 (3.9%) | 21 (4.8%) | 50 (5.1%) | | | Lymphocytes (×109/L) | Median [Q1, Q3] | 1.85 [1.51, 2.23] | 1.83 [1.52, 2.23] | 1.85 [1.49, 2.22] | | Missing, n (%) | 42 (4.1%) | 16 (3.6%) | 56 (5.7%) | | | Monocytes (×109/L) | Median [Q1, Q3] | 0.37 [0.31, 0.45] | 0.38 [0.31, 0.47] | 0.38 [0.31, 0.45] | | Missing, n (%) | 47 (4.6%) | 19 (4.3%) | 68 (6.9%) | | | Platelets (×109/L) | Median [Q1, Q3] | 243.00 [205.00, 285.00] | 241.00 [203.00, 282.00] | 241.00 [204.00, 281.00] | | Missing, n (%) | 40 (3.9%) | 16 (3.6%) | 33 (3.4%) | | | Endometriosis status | No endometriosis | 677 (65.8%) | 290 (65.8%) | 646 (65.8%) | | Endometriosis | 352 (34.2%) | 151 (34.2%) | 336 (34.2%) | | | Missing, n (%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | | | Previous pregnancy history | Previous pregnancy: yes | 454 (46.9%) | 200 (47.5%) | 441 (47.4%) | | Previous pregnancy: no | 515 (53.1%) | 221 (52.5%) | 490 (52.6%) | | | Missing, n (%) | 60 (5.8%) | 20 (4.5%) | 51 (5.2%) | | | Surgical status | No | 155 (15.1%) | 71 (16.1%) | 141 (14.4%) | | Yes | 874 (84.9%) | 370 (83.9%) | 841 (85.6%) | | | Missing, n (%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | | | Ovarian endometrioma | No | 899 (87.4%) | 382 (86.6%) | 869 (88.5%) | | Yes | 130 (12.6%) | 59 (13.4%) | 113 (11.5%) | | | Missing, n (%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | | | TVUS suspicion of endometriosis | No | 864 (87.2%) | 363 (86.8%) | 805 (86.8%) | | Yes | 127 (12.8%) | 55 (13.2%) | 122 (13.2%) | | | Missing, n (%) | 38 (3.7%) | 23 (5.2%) | 55 (5.6%) | | | Dysmenorrhea | No | 448 (43.7%) | 208 (47.2%) | 409 (41.8%) | | Yes | 577 (56.3%) | 233 (52.8%) | 569 (58.2%) | | | Missing, n (%) | 4 (0.4%) | 0 (0.0%) | 4 (0.4%) | | | Chronic pelvic pain | No | 593 (57.7%) | 257 (58.3%) | 572 (58.4%) | | Yes | 435 (42.3%) | 184 (41.7%) | 407 (41.6%) | | | Missing, n (%) | 1 (0.1%) | 0 (0.0%) | 3 (0.3%) | | | Infertility | No | 728 (71.4%) | 309 (70.5%) | 705 (72.4%) | | Yes | 292 (28.6%) | 129 (29.5%) | 269 (27.6%) | | | Missing, n (%) | 9 (0.9%) | 3 (0.7%) | 8 (0.8%) | Baseline characteristics by study partition. Continuous variables are presented as median (Q1–Q3) using observed, nonmissing measurements. Categorical variables are presented as n (%), with percentages calculated using nonmissing observations as the denominator. Missing-data percentages were calculated using the total number of participants in each study partition. All comparisons are descriptive. CA125, cancer antigen 125; EM, endometriosis; HDL, high-density lipoprotein cholesterol; LDL, low-density lipoprotein cholesterol; TVUS, transvaginal ultrasonography. 3.2 CA125 distribution and missing-data assessment Among participants with observed CA125 measurements, concentrations were higher in patients with EM (n = 776; median, 35.60 U/mL; IQR, 21.78–61.03) than in disease controls (n = 1,490; median, 17.05 U/mL; IQR, 11.39–26.09), although the distributions overlapped. CA125 measurements were missing in 63 patients with EM and 123 disease controls. These descriptive statistics were calculated using observed measurements only, without substitution of imputed values. A diagnostic reconstruction of the fivefold multiple imputation by chained equations (MICE) procedure was conducted to examine the distribution of imputed CA125 values among participants with missing measurements. Across the reconstructed datasets, imputed values tended to be lower than observed values among patients with EM and slightly higher among disease controls, resulting in greater overlap between groups (Supplementary Table S8). These findings are descriptive and should be interpreted cautiously because the reconstructed imputations were not verified exports of the original model-development imputation chains. Therefore, they cannot establish the exact values imputed during the original analysis or their effects on individual model predictions. 3.3 Feature-selection stability and development-stage model comparison Feature-selection stability was evaluated across 50 outer resampling iterations using least absolute shrinkage and selection operator (LASSO) regression and Boruta. Among the 12 candidate predictors, CA125, high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), and neutrophil count (NEUT) were consistently retained by the intersection of the two selection methods, each achieving a selection frequency of 100%. The remaining predictors showed lower intersection frequencies, including monocyte count (MONO; 48%), triglycerides (TG; 46%), platelet count (PLT; 40%), lymphocyte count (LYM; 36%), and total cholesterol (TC; 6%). Age and body mass index (BMI) were not retained by the intersection in any of the 50 iterations. White blood cell count (WBC) was excluded from the eligible predictor set in all iterations because its absolute Pearson correlation with neutrophil count (NEUT) exceeded 0.90; therefore, its zero selection frequency reflected correlation-based exclusion rather than nonselection by LASSO or Boruta (Figure 2A; Supplementary Table S1). Thus, four predictors exceeded the descriptive stability threshold of 70%. The stability assessment was distinct from the final feature-selection procedure. When LASSO (λ.1se) and Boruta were applied to the five multiply imputed versions of the complete training dataset, eight predictors were retained through majority voting across imputations: TG, HDL, LDL, LYM, NEUT, MONO, PLT, and CA125 (Figure 2B). These eight variables constituted the predictor set for the final model. Six machine learning algorithms were subsequently compared using five repeats of nested 10-fold cross-validation within the training dataset. The gradient boosting machine (GBM) achieved the highest mean outer cross-validation area under the receiver operating characteristic curve (AUC; 0.8483), followed closely by extreme gradient boosting (XGBoost; 0.8470), adaptive boosting (AdaBoost; 0.8397), neural network (0.8351), support vector machine (0.8275), and logistic regression (0.8113) (Figure 3). The corresponding standard deviations across the five cross-validation repeats were 0.0036, 0.0036, 0.0021, 0.0021, 0.0062, and 0.0053, respectively. Based on the prespecified development-stage selection criterion of the highest mean cross-validation AUC, GBM was selected as the final modeling algorithm. The difference in mean AUC between GBM and XGBoost was small (approximately 0.0013), and no statistically significant superiority of GBM over XGBoost was established. Algorithm selection was completed without reference to performance in the internal test or temporal validation cohorts. 3.4 Independent validation and comparison with clinical reference models The final eight-predictor GBM model was evaluated in the held-out internal test set (n = 441) and the independent temporal validation cohort (n = 982). The model achieved an AUC of 0.8401 (95% CI, 0.8006–0.8797) in the internal test set and 0.8565 (95% CI, 0.8320–0.8810) in the temporal validation cohort (Figure 4; Table 2). Table 2 | Cohort | Model | AUC (95% CI) | Brier score | Sensitivity (%) | Specificity (%) | PPV (%) | NPV (%) | ΔAUC vs. GBM | Paired DeLong P | |---|---|---|---|---|---|---|---|---|---| | Internal test (n = 441) | CA125 alone | 0.735 (0.684–0.787) | 0.187 | 33.8 | 92.8 | 70.8 | 72.9 | 0.105 | <0.001 | | Basic clinical model | 0.768 (0.722–0.814) | 0.182 | 34.4 | 93.4 | 73.2 | 73.2 | 0.072 | <0.001 | | | Extended clinical model | 0.833 (0.791–0.875) | 0.146 | 53.6 | 92.4 | 78.6 | 79.3 | 0.008 | 0.758 | | | GBM (8 predictors) | 0.840 (0.801–0.880) | 0.147 | 55.6 | 91.4 | 77.1 | 79.8 | Reference | – | | | Temporal validation (n = 982) | CA125 alone | 0.758 (0.726–0.790) | 0.19 | 31.8 | 92.6 | 69 | 72.3 | 0.098 | <0.001 | | Basic clinical model | 0.768 (0.738–0.799) | 0.185 | 33 | 93.2 | 71.6 | 72.8 | 0.088 | <0.001 | | | Extended clinical model | 0.841 (0.815–0.867) | 0.148 | 57.4 | 89 | 73.1 | 80.1 | 0.016 | 0.294 | | | GBM (8 predictors) | 0.856 (0.832–0.881) | 0.142 | 59.5 | 88.9 | 73.5 | 80.8 | Reference | – | Independent validation performance of the GBM model and clinical comparators. All performance estimates were obtained from independent validation predictions. Classification metrics were calculated at the fixed probability threshold of 0.50. ΔAUC was calculated as the AUC of the GBM model minus that of the comparator, and P values were obtained using paired DeLong tests. The basic clinical model included CA125, age, body mass index, and previous pregnancy history. The extended clinical model additionally included dysmenorrhea, chronic pelvic pain, infertility, and transvaginal ultrasonography findings suggestive of endometriosis. Positive and negative predictive values reflect the observed case composition of each evaluation cohort and should not be interpreted as population-level estimates. AUC, area under the receiver operating characteristic curve; CI, confidence interval; GBM, gradient boosting machine; NPV, negative predictive value; PPV, positive predictive value. At the fixed classification threshold of 0.50, sensitivity and specificity were 55.63% and 91.38%, respectively, in the internal test set. Corresponding values in the temporal validation cohort were 59.52% and 88.85%. These findings indicated relatively high specificity but more limited sensitivity at the selected threshold. In the internal test set, the AUCs of CA125 alone, the basic clinical model, and the extended clinical model were 0.7353, 0.7684, and 0.8326, respectively. Corresponding AUCs in the temporal validation cohort were 0.7583, 0.7683, and 0.8410. The GBM achieved numerically higher AUCs than all three reference models in both evaluation cohorts (Figure 4; Table 2). However, the improvement over the extended clinical model was modest and not statistically significant. The difference in AUC between the GBM and the extended clinical model was 0.0075 in the internal test set (DeLong P = 0.758) and 0.0155 in the temporal validation cohort (DeLong P = 0.294). Therefore, although the GBM demonstrated consistent discrimination across the two independent evaluation cohorts, these results did not establish superior discrimination relative to the extended clinical model. 3.5 Calibration and decision curve analysis The calibration performance of the final GBM model was assessed separately in the internal test set and temporal validation cohort (Figures 5A,B). The Brier scores were 0.1471 and 0.1421, respectively. The calibration intercept was 0.0754 in the internal test set and 0.0137 in the temporal validation cohort, while the corresponding calibration slopes were 1.0889 and 1.1295. These results indicated relatively small overall calibration deviations, although the slopes exceeding 1.0 suggested some departure from ideal calibration in both cohorts. Decision curve analysis (DCA) was performed to explore the potential net benefit of the GBM model across a range of threshold probabilities using the original, unrecalibrated GBM predictions (Figures 5C,D). The analyses were based on the observed EM case proportions in the internal test set and temporal validation cohort, without adjustment for the prevalence of a defined target population. Accordingly, the decision curves provide an exploratory assessment of cohort-specific net benefit rather than definitive evidence of clinical utility in routine gynecological practice. Taken together, the calibration and decision curve analyses provided complementary information beyond discrimination alone. However, prospective evaluation in representative clinical populations is required to establish whether the model can improve diagnostic decision-making. 3.6 Subgroup and sensitivity analyses Subgroup analyses were conducted to explore the discrimination of the final GBM model across clinically relevant patient subsets (Table 3). When the evaluation was restricted to surgically evaluated disease controls, the AUC was 0.8411 in the internal test set and 0.8609 in the temporal validation cohort, closely resembling the corresponding results in the overall evaluation populations. Of the disease controls, 522/677 in training, 219/290 in internal testing, and 505/646 in temporal validation had a surgical evaluation recorded. Table 3 | Cohort | Subgroup | Total (n) | EM (n) | NEM (n) | AUC (95% CI) | Sensitivity (%) | Specificity (%) | Brier score | |---|---|---|---|---|---|---|---|---| | Internal test | Overall validation cohort | 441 | 151 | 290 | 0.840 (0.801–0.880) | 55.6 | 91.4 | 0.147 | | Surgically evaluated NEM controls | 370 | 151 | 219 | 0.841 (0.800–0.882) | 55.6 | 91.8 | 0.159 | | | rASRM stage I–II | 286 | 67 | 219 | 0.814 (0.754–0.874) | 50.7 | 91.8 | 0.134 | | | rASRM stage III–IV | 303 | 84 | 219 | 0.863 (0.814–0.911) | 59.5 | 91.8 | 0.125 | | | EM without endometrioma | 311 | 92 | 219 | 0.815 (0.764–0.865) | 47.8 | 91.8 | 0.151 | | | EM with endometrioma | 278 | 59 | 219 | 0.882 (0.824–0.940) | 67.8 | 91.8 | 0.104 | | | TVUS-negative patients | 363 | 95 | 268 | 0.847 (0.801–0.894) | 56.8 | 91.4 | 0.128 | | | TVUS-positive patients (exploratory) | 55 | 49 | 6 | Not estimated | 57.1 | 83.3 | 0.264 | | | Temporal validation | Overall validation cohort | 982 | 336 | 646 | 0.856 (0.832–0.881) | 59.5 | 88.9 | 0.142 | | Surgically evaluated NEM controls | 841 | 336 | 505 | 0.861 (0.836–0.886) | 59.5 | 89.7 | 0.15 | | | rASRM stage I–II | 648 | 143 | 505 | 0.829 (0.790–0.868) | 51 | 89.7 | 0.127 | | | rASRM stage III–IV | 698 | 193 | 505 | 0.885 (0.858–0.912) | 65.8 | 89.7 | 0.119 | | | EM without endometrioma | 728 | 223 | 505 | 0.843 (0.812–0.874) | 55.2 | 89.7 | 0.142 | | | EM with endometrioma | 618 | 113 | 505 | 0.896 (0.864–0.929) | 68.1 | 89.7 | 0.101 | | | TVUS-negative patients | 805 | 212 | 593 | 0.842 (0.811–0.873) | 56.6 | 89.2 | 0.132 | | | TVUS-positive patients (exploratory) | 122 | 103 | 19 | 0.815 (0.698–0.931) | 65 | 73.7 | 0.213 | Subgroup performance of the locked GBM model. Subgroup analyses used the locked GBM predictions without model refitting, predictor reselection, recalibration, or threshold optimization. Sensitivity and specificity were calculated at the fixed probability threshold of 0.50. For analyses stratified by revised American Society for Reproductive Medicine (rASRM) stage and ovarian endometrioma status, the corresponding endometriosis cases were compared against the same surgically evaluated disease-control population within each cohort. Consequently, subgroup denominators overlap and should not be summed. Analyses stratified by transvaginal ultrasonography (TVUS) findings were restricted to participants with the corresponding ultrasound status. The internal-test TVUS-positive subgroup included only six disease controls; its AUC and confidence interval were therefore not estimated. All subgroup analyses were exploratory, and between-subgroup differences were not formally tested. Surgical evaluation alone does not definitively exclude occult endometriosis in every disease control. AUC confidence intervals were calculated using the DeLong method where estimable. Discrimination varied descriptively according to EM severity. Among patients with revised American Society for Reproductive Medicine (rASRM) stage I–II disease, the AUCs were 0.8141 and 0.8286 in the internal test and temporal validation cohorts, respectively. The corresponding AUCs for stage III–IV disease were 0.8625 and 0.8849. A similar pattern was observed according to ovarian endometrioma status. Among patients without ovarian endometrioma, the AUCs were 0.8147 and 0.8430 in the internal test and temporal validation cohorts, respectively. Among those with ovarian endometrioma, the corresponding AUCs were 0.8822 and 0.8963. These subgroup findings were exploratory, and the observed differences were not subjected to formal interaction testing. To assess sensitivity to missing-data handling, an alternative GBM model was refitted using median imputation instead of the primary MICE-based approach. The median-imputation model achieved AUCs of 0.8406 and 0.8575 in the internal test and temporal validation cohorts, respectively, compared with 0.8401 and 0.8565 for the primary model. The small differences suggested that overall discrimination was relatively insensitive to the alternative imputation strategy in this analysis (Supplementary Table S3). In an exploratory age-adjusted analysis of 2,305 participants with complete age and lipid measurements, HDL (OR per SD, 0.382; 95% CI, 0.339–0.430; P < 0.001) and LDL (OR per SD, 0.649; 95% CI, 0.588–0.717; P < 0.001) remained associated with EM, whereas TG did not (OR per SD, 0.967; 95% CI, 0.883–1.059; P = 0.469). Findings in the training partition were directionally consistent (Supplementary Table S7). These exploratory associations do not establish causality or eliminate residual confounding. 3.7 Complete-predictor performance and computational reproducibility Among participants with complete measurements of all eight selected predictors, 331 individuals in the internal test set (120 EM cases and 211 disease controls) and 656 individuals in the temporal validation cohort (239 EM cases and 417 disease controls) were eligible for complete-predictor evaluation. The frozen GBM prediction pipeline achieved an AUC of 0.8535 (95% CI, 0.8110–0.8960) in the internal test set and 0.8559 (95% CI, 0.8263–0.8855) in the temporal validation cohort. The corresponding Brier scores were 0.1458 and 0.1479, calibration intercepts were 0.1465 and 0.1354, and calibration slopes were 1.1308 and 1.0743, respectively. At the fixed classification threshold of 0.50, sensitivity was 56.67% and 59.00%, whereas specificity was 91.00% and 88.49%, respectively. To assess numerical reproducibility, predictions generated using the publicly released GBM models and frozen preprocessing parameters were compared with the corresponding stored model predictions for all 987 participants with complete measurements of the eight selected predictors across the two evaluation cohorts. The maximum absolute difference in predicted probabilities was approximately 5.0 × 10−16, indicating agreement within machine precision. Detailed performance and numerical reproducibility results are presented in Supplementary Table S5. These analyses demonstrated that the frozen GBM pipeline could reproduce predictions for participants with complete measurements of all eight predictors. However, the standalone prediction function does not implement MICE for new observations, and these results do not establish equivalent performance among patients with missing predictors or generalizability to independent clinical populations. 4 Discussion In this retrospective diagnostic modeling study, we developed and temporally validated a gradient boosting machine (GBM) model integrating eight routinely available laboratory indicators to identify existing endometriosis (EM) among women undergoing gynecological evaluation. The final model incorporated triglycerides (TG), high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), lymphocyte count (LYM), neutrophil count (NEUT), monocyte count (MONO), platelet count (PLT), and CA125. Among six candidate algorithms, GBM achieved the highest mean area under the receiver operating characteristic curve (AUC) during repeated nested cross-validation. The final model demonstrated consistent discrimination in the held-out internal test set and temporal validation cohort, with AUCs of 0.840 and 0.856, respectively, although calibration slopes indicated some departures from ideal calibration. Although the GBM achieved numerically higher AUCs than CA125 alone and the clinical reference models, its improvement over the extended clinical model was modest and not statistically significant. These findings suggest that routinely measured hematological and lipid indicators may provide complementary diagnostic information, while emphasizing the importance of evaluating machine learning models against clinically relevant reference approaches. CA125 remains one of the most extensively investigated circulating biomarkers for EM, although its limited sensitivity restricts its use as a standalone diagnostic test ( , , ). In the present study, CA125 alone yielded AUCs of 0.735 and 0.758 in the internal test and temporal validation cohorts, respectively, compared with 0.840 and 0.856 for the multivariable GBM model. These findings support the potential value of integrating CA125 with additional laboratory information. However, the observed improvement cannot be attributed to any individual predictor, and the present analysis does not establish the independent incremental contribution of each laboratory variable. Previous biomarker reviews have also emphasized the heterogeneity of candidate markers across endometriosis phenotypes and biological compartments ( ). The proposed model should therefore be regarded as a potential complement to CA125 and established clinical assessment rather than a replacement for either.The inclusion of TG, HDL, and LDL suggests that conventional lipid measurements may contribute predictive information when considered jointly with hematological indicators and CA125. Previous studies have reported differences in circulating lipid profiles among women with EM, although the direction and magnitude of associations have varied across populations ( ). Lipidomic and metabolomic investigations have also identified disease-associated metabolic signatures ( , ). Nevertheless, conventional lipid measurements are influenced by nutritional status, metabolic conditions, hormonal exposure, and medication use. Their retention in a prediction model does not establish independent associations with EM or direct involvement in disease pathogenesis. Accordingly, these lipid parameters should be interpreted as components of a multivariable diagnostic profile rather than disease-specific biomarkers or evidence of a causal metabolic mechanism.Similarly, the retention of LYM, NEUT, MONO, and PLT suggests that routinely measured hematological parameters may provide complementary information for distinguishing EM from other gynecological conditions. EM involves complex inflammatory and immune processes, and previous studies have implicated neutrophils, macrophages, lymphocytes, and other immune cell populations in disease-associated responses ( , ). However, peripheral blood cell counts are nonspecific and cannot directly characterize immune-cell activity within endometriotic lesions. Notably, NEUT was consistently retained by the intersection of LASSO and Boruta across all 50 resampling iterations, whereas MONO, LYM, and PLT showed lower selection frequencies in the descriptive stability analysis. All four variables were subsequently retained through the separate final feature-selection procedure. This distinction illustrates that predictor retention depends on the specified selection strategy and the joint predictive information provided by candidate variables, rather than demonstrating a direct mechanistic role.An important methodological feature of this study was the separation of model development from independent evaluation. Six candidate algorithms were compared using repeated nested cross-validation within the training dataset, and GBM was selected according to the highest mean development-stage AUC. Its performance was nearly identical to that of extreme gradient boosting (XGBoost), and the findings do not establish the superiority of one algorithm over the other. The selected GBM was subsequently evaluated in a held-out internal test set and a temporally independent cohort that had not been used for algorithm selection. This design reduced the risk of optimistic performance estimates arising from repeated model evaluation on the same data ( , ). Nevertheless, temporal validation within a single institution does not establish geographic transportability, because both evaluation cohorts may share referral patterns, diagnostic practices, laboratory procedures, and patient characteristics. Independent multicenter validation remains necessary ( ).Comparison with clinically relevant reference models is essential for assessing the potential added value of a multivariable prediction approach. The GBM demonstrated higher discrimination than CA125 alone and the basic clinical model incorporating CA125, age, body mass index, and previous pregnancy history. However, its advantage was substantially smaller when compared with the extended clinical model, which additionally incorporated dysmenorrhea, chronic pelvic pain, infertility, and ultrasonographic findings suggestive of EM. The differences in AUC between GBM and the extended clinical model were 0.0075 in the internal test set and 0.0155 in the temporal validation cohort, neither of which was statistically significant. Thus, although the laboratory-based GBM achieved numerically similar discrimination to the extended clinical reference model, the present findings do not establish equivalence, superior diagnostic accuracy, or sufficient incremental clinical value to justify replacing symptom assessment and imaging. The operating characteristics of the model further inform its potential clinical role. At the fixed probability threshold of 0.50, specificity was 91.38% in the internal test set and 88.85% in the temporal validation cohort, whereas sensitivity was 55.63% and 59.52%, respectively. Consequently, a substantial proportion of patients with EM would not be identified at this threshold. A negative model result should not be used to exclude EM, particularly when symptoms, examination findings, or imaging raise clinical suspicion. The model is therefore better considered an investigational adjunct to established gynecological assessment rather than a standalone diagnostic or screening tool. Its operating threshold and consequences for downstream clinical decisions require prospective evaluation in a clearly defined target population. Calibration and clinical utility are important considerations beyond discrimination alone. Although the model showed relatively small calibration intercept deviations, calibration slopes exceeded 1.0 in both independent evaluation cohorts, indicating departures from ideal calibration and suggesting that the predicted risks were less dispersed than the observed outcomes. Calibration should therefore be evaluated independently of discrimination rather than inferred from AUC alone ( ). Decision curve analysis (DCA) provided an exploratory assessment of potential net benefit across different thresholds applied to the original GBM prediction scores ( , ). However, these analyses were based on the observed case proportions in the retrospective evaluation cohorts, which should not be interpreted as the prevalence of endometriosis in an unselected clinical population. Moreover, because the GBM scores were not recalibrated to the baseline disease probability of a defined target population, the thresholds used in DCA should not be regarded as transportable absolute-risk thresholds. Consequently, the observed decision curves do not establish clinical benefit in other settings without further assessment of population-specific calibration, clinically appropriate decision thresholds, and potential harms associated with missed cases or unnecessary investigations.The exploratory subgroup analyses suggested that discrimination may vary across clinical presentations. AUCs were numerically higher among patients with revised American Society for Reproductive Medicine (rASRM) stage III–IV disease than among those with stage I–II disease, and among patients with ovarian endometrioma than among those without endometrioma. These patterns may reflect differences in disease presentation or the extent to which laboratory indicators distinguish particular case groups from disease controls. However, subgroup comparisons were descriptive, involved potentially overlapping disease-control populations, and were not supported by formal interaction testing. The numerically lower AUCs in earlier-stage disease and in patients without ovarian endometrioma warrant further investigation, given the diagnostic challenges associated with these presentations. The sensitivity analysis using median imputation yielded AUCs similar to those obtained using the primary multiple-imputation approach, suggesting that discrimination was not materially altered by the alternative missing-data strategy in the present analysis. Nevertheless, this agreement does not eliminate potential bias arising from missing data or establish the validity of the assumptions underlying multiple imputation ( ). Further evaluation in datasets with more complete predictor ascertainment is warranted.To facilitate reproducibility and independent evaluation, we publicly released the five fitted GBM models, frozen training-derived preprocessing parameters, and an executable prediction function. In the complete-predictor subsets, the frozen prediction pipeline achieved AUCs of 0.8535 (n = 331) and 0.8559 (n = 656) in the internal test and temporal validation cohorts, respectively. Across 987 participants with complete measurements of all eight predictors, predictions generated using the released model and preprocessing parameters agreed with the corresponding stored predictions to numerical precision (maximum absolute difference, approximately 5.0 × 10−16). The standalone prediction function requires complete measurements of all eight predictors and does not perform multiple imputation by chained equations (MICE). Accordingly, these reproducibility results apply to complete-predictor inputs and should not be interpreted as evidence of validated prediction for patients with missing predictors or of generalizability to independent clinical populations. Several limitations should be acknowledged. First, the retrospective, single-center design and use of disease controls from a gynecological care population may limit generalizability. Second, diagnostic verification was not uniform: EM cases were surgically and histologically confirmed, whereas disease controls were evaluated through different clinical pathways. This may introduce verification-related bias or outcome misclassification. Third, the retrospective data did not permit comprehensive adjustment for factors potentially affecting lipid and hematological measurements, including menstrual-cycle timing, hormonal exposure, metabolic conditions, and other clinical characteristics. Fourth, the complete-predictor validation subsets represented only a proportion of the independent evaluation cohorts, introducing potential selection effects associated with predictor completeness. Fifth, although temporal validation was performed, multicenter external validation and prospective clinical-impact assessment remain necessary. The importance of evaluating bias and applicability when interpreting clinical prediction models is also emphasized in contemporary methodological guidance ( ). Finally, subgroup analyses were exploratory, and the selected predictors should not be interpreted as evidence of causal or mechanistic relationships with EM. The observed case proportions may also differ from those in the intended deployment population, limiting direct interpretation of predicted probabilities and decision-curve findings in routine practice.Despite these limitations, the study provides a structured evaluation of a diagnostic model integrating CA125, conventional lipid measurements, and hematological parameters, incorporating repeated nested cross-validation, held-out testing, temporal validation, calibration assessment, clinical reference comparisons, sensitivity analyses, and public release of fitted model objects. However, the absence of a statistically significant improvement over the extended clinical model and the limited sensitivity at the evaluated threshold indicate that further validation and assessment of clinical utility are required before clinical implementation. 5 Conclusion This study developed and temporally validated an eight-predictor gradient boosting machine (GBM) model based on routinely available laboratory indicators for identifying endometriosis among women undergoing gynecological evaluation. The model demonstrated consistent discrimination in the internal test and temporal validation cohorts (AUCs of 0.840 and 0.856, respectively). However, it did not significantly outperform an extended clinical model incorporating symptoms and transvaginal ultrasonography, and its incremental clinical value remains uncertain. The publicly available fitted model and prediction pipeline facilitate reproducibility and independent evaluation. These findings support further investigation of laboratory-based multivariable models as potential adjuncts to clinical assessment rather than replacements for established diagnostic approaches. Multicenter external validation and prospective studies are warranted to establish generalizability and determine whether the model provides meaningful clinical benefit. Statements Data availability statement The data analyzed in this study is subject to the following licenses/restrictions: the data are not publicly available due to privacy and ethical restrictions. De-identified data may be available from the corresponding author upon reasonable request, subject to institutional ethics approval. The fitted ensemble of five gradient boosting machine (GBM) models, frozen training-derived winsorization and standardization parameters, executable R prediction function, example inputs, and supporting documentation are publicly available on Zenodo (Version 2.0.0; DOI: 10.5281/zenodo.23408524). The prediction function requires complete measurements of all eight selected predictors and does not perform missing-data imputation. Multiple imputation by chained equations (MICE) was used in the retrospective model-development and validation analyses, with the corresponding methodological reference code provided in the repository. Requests to access these datasets should be directed to Guoliu Ye, [email protected]. Ethics statement This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Bengbu Medical University (Approval No. [2023]449) and was conducted in accordance with the Declaration of Helsinki. The requirement for individual informed consent was waived by the Ethics Committee because of the retrospective nature of the study. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because this work is a retrospective study. Author contributions QH: Supervision, Software, Formal Analysis, Writing – review & editing, Conceptualization, Data curation, Project administration, Writing – original draft, Methodology, Investigation, Visualization, Validation, Resources. HL: Writing – original draft, Software, Methodology, Data curation, Visualization. SM: Software, Writing – original draft, Investigation, Methodology, Data curation, Visualization. ZZ: Visualization, Software, Writing – original draft, Data curation, Validation. TR: Data curation, Writing – original draft, Formal Analysis, Validation, Methodology. LW: Resources, Methodology, Formal Analysis, Writing – original draft, Data curation. JW: Formal Analysis, Writing – original draft, Methodology, Investigation. GY: Supervision, Funding acquisition, Writing – review & editing, Project administration, Resources, Formal Analysis. Funding The author(s) declared that financial support was received for this work and/or its publication. Funding for this work was provided by the Longhu Talent Project of Bengbu Medical University (grant number: LH250203007). Acknowledgments We extend our sincere gratitude to all the study participants for their invaluable contribution. Conflict of interest The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Generative AI statement The author(s) declared that generative AI was not used in the creation of this manuscript. Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us. Publisher’s note All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Supplementary material The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmed.2026.1934454/full#supplementary-material Abbreviations AdaBoost, adaptive boosting; AUC, area under the receiver operating characteristic curve; BMI, body mass index; CA125, cancer antigen 125; CI, confidence interval; DCA, decision curve analysis; EM, endometriosis; GBM, gradient boosting machine; HDL, high-density lipoprotein cholesterol; IQR, interquartile range; LASSO, least absolute shrinkage and selection operator; LDL, low-density lipoprotein cholesterol; LR, logistic regression; LYM, lymphocyte count; MICE, multiple imputation by chained equations; MONO, monocyte count; NEUT, neutrophil count; NN, neural network; NPV, negative predictive value; OR, odds ratio; PLT, platelet count; PPV, positive predictive value; rASRM, Revised American Society for Reproductive Medicine; ROC, receiver operating characteristic; SD, standard deviation; SMD, standardized mean difference; SVM, support vector machine; TC, total cholesterol; TG, triglycerides; TRIPOD + AI, transparent reporting of a multivariable prediction model for individual prognosis or diagnosis–artificial intelligence; TVUS, transvaginal ultrasonography; WBC, white blood cell count; XGBoost, extreme gradient boosting.

References

1. ZondervanKTBeckerCMMissmerSA. Endometriosis. N Engl J Med. (2020) 382:1244–56. 10.1056/NEJMra1810764 2. As-SanieSMackenzieSCMorrisonLSchrepfAZondervanKTHorneAWet al. Endometriosis: a review. J Am Med Assoc. (2025) 334:64–78. 10.1001/jama.2025.2975 3. SimoensSDunselmanGDirksenCHummelshojLBokorABrandesIet al. The burden of endometriosis: costs and quality of life of women with endometriosis and treated in referral centres. Hum Reprod. (2012) 27:1292–9. 10.1093/humrep/des073 4. De CortePKlinghardtMVon StockumSHeinemannK. Time to diagnose endometriosis: current status, challenges and regional characteristics—a systematic literature review. BJOG. (2025) 132:118–30. 10.1111/1471-0528.17973 5. FryerJMason-JonesAJWoodwardA. Understanding diagnostic delay for endometriosis: a scoping review using the social-ecological framework. Health Care Women Int. (2025) 46:335–51. 10.1080/07399332.2024.2413056 6. BeckerCMBokorAHeikinheimoOHorneAJansenFKieselLet al. ESHRE guideline: endometriosis. Hum Reprod Open. (2022) 2022. 10.1093/hropen/hoac009 7. AveryJCDeslandesAFregerSMLeonardiMLoGCarneiroGet al. Noninvasive diagnostic imaging for endometriosis part 1: a systematic review of recent developments in ultrasound, combination imaging, and artificial intelligence. Fertil Steril. (2024) 121:164–88. 10.1016/j.fertnstert.2023.12.008 8. HirschMDuffyJMNDavisCJNieves PlanaMKhanKS. Diagnostic accuracy of cancer antigen 125 for endometriosis: a systematic review and meta-analysis. BJOG. (2016) 123:1761–8. 10.1111/1471-0528.14055 9. NisenblatVBossuytPMMShaikhRFarquharCJordanVScheffersCSet al. Blood biomarkers for the non-invasive diagnosis of endometriosis. Cochrane Database Syst Rev. (2016) 2016. 10.1002/14651858.CD012179 10. BrulportABourdonMVaimanDDrouetCPocate-CherietKBouzidKet al. An integrated multi-tissue approach for endometriosis candidate biomarkers: a systematic review. Reprod Biol Endocrinol. (2024) 22:21. 10.1186/s12958-023-01181-8 11. Zare MoghaddamMAnsariniyaHSeifatiSMZareFFesahatF. Immunopathogenesis of endometriosis: an overview of the role of innate and adaptive immune cells and their mediators. Am J Reprod Immunol. (2022) 87. 10.1111/aji.13537 12. WangXJiaYLiDGuoXZhouZQiMet al. The abundance and function of neutrophils in the endometriosis systemic and pelvic microenvironment. Mediators Inflamm. (2023) 2023:1481489. 10.1155/2023/1481489 13. MeloASRosa-e-SilvaJCRosa-e-SilvaACJSPoli-NetoOBFerrianiRAVieiraCS. Unfavorable lipid profile in women with endometriosis. Fertil Steril. (2010) 93:2433–6. 10.1016/j.fertnstert.2009.08.043 14. StarodubtsevaNChagovetsVBorisovaASalimovaDAleksandrovaNChinginKet al. Identification of potential endometriosis biomarkers in peritoneal fluid and blood plasma via shotgun lipidomics. Clin Mass Spectrom. (2019) 13:21–6. 10.1016/j.clinms.2019.05.007 15. TomkinsNEGirlingJEBoughtonBHoldsworth-CarsonSJ. Is there a role for small molecule metabolite biomarkers in the development of a diagnostic test for endometriosis?Syst Biol Reprod Med. (2022) 68:89–112. 10.1080/19396368.2022.2027045 16. ZhaoNHaoTZhangFNiQZhuDWangYet al. Application of machine learning techniques in the diagnosis of endometriosis. BMC Womens Health. (2024) 24:491. 10.1186/s12905-024-03334-2 17. DorienFOFassbenderAVan BreeRLaenenAPeterseDPVanhieAet al. Technical verification and assessment of independent validation of biomarker models for endometriosis. Biomed Res Int. (2019) 2019:3673060. 10.1155/2019/3673060 18. ZhangBLvXLiDZhangLRuZMaY. Diagnostic accuracy of machine learning for endometriosis: a systematic review and meta-analysis. Front Endocrinol. (2026) 16:1735567. 10.3389/fendo.2025.1735567 19. CollinsGSDhimanPMaJSchlusselMMArcherLVan CalsterBet al. Evaluation of clinical prediction models (part 1): from development to external validation. Br Med J. (2024) 384:e074819. 10.1136/bmj-2023-074819 20. RileyRDArcherLSnellKIEEnsorJDhimanPMartinGPet al. Evaluation of clinical prediction models (part 2): how to undertake an external validation study. Br Med J. (2024) 384:e074820. 10.1136/bmj-2023-074820 21. Van CalsterBMcLernonDJVan SmedenMWynantsLSteyerbergEW. Calibration: the Achilles heel of predictive analytics. BMC Med. (2019) 17:230. 10.1186/s12916-019-1466-7 22. SteyerbergEWVergouweY. Towards better clinical prediction models: seven steps for development and an ABCD for validation. Eur Heart J. (2014) 35:1925–31. 10.1093/eurheartj/ehu207 23. VickersAJElkinEB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. (2006) 26:565–74. 10.1177/0272989X06295361 24. VickersAJVan CalsterBSteyerbergEW. Net benefit approaches to the evaluation of prediction models, molecular markers, and diagnostic tests. Br Med J. (2016) 352:i6. 10.1136/bmj.i6 25. SterneJACWhiteIRCarlinJBSprattMRoystonPKenwardMGet al. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. Br Med J. (2009) 338:b2393. 10.1136/bmj.b2393 26. MoonsKGMDamenJAAKaulTHooftLAndaur NavarroCDhimanPet al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. Br Med J. (2025) 388:e082505. 10.1136/bmj-2024-082505 27. CollinsGSMoonsKGMDhimanPRileyRDBeamALVan CalsterBet al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. Br Med J. (2024) 385:e078378. 10.1136/bmj-2023-078378 Summary

Keywords

diagnostic model, endometriosis, gradient boosting machine, machine learning, routine laboratory indicators, temporal validation Citation Hu Q, Li H, Ma S, Zhao Z, Rong T, Wang L, Wei J and Ye G (2026) Development and temporal validation of a gradient boosting machine model for endometriosis using routine laboratory indicators. Front. Med. 13:1934454. doi: 10.3389/fmed.2026.1934454 Received 10 July 2026 Revised 17 October 2026 Accepted 08 November 2026 Published 18 November 2026 Volume 13 - 2026 Edited by Pablo C. Ortiz-Lazareno, Centro de Investigación Biomédica de Occidente (CIBO), Mexico Reviewed by Alejandro Martinez-Martinez, Universidad Autónoma de Ciudad Juárez, Mexico Hiroshi Koike, The University of Tokyo, Japan Updates Copyright © 2026 Hu, Li, Ma, Zhao, Rong, Wang, Wei and Ye. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. *Correspondence: Guoliu Ye [email protected] Disclaimer All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Text is read by the "Ask this paper" AI Q&A widget below. Extraction quality varies by source — PMC NXML preserves structure cleanly, OA-HTML may include some navigation residue, and OA-PDF can have broken hyphenation. The publisher copy (via DOI) is the canonical version.

My notes (saved in your browser only)

⚙ Ask this paper AI returns verbatim quotes from the full text · source: oa-doi-fallback ⓘ

Answers must be backed by verbatim quotes from this paper's full text. Hallucinated quotes are dropped automatically; if no verbatim passage answers the question, we say so. How this works

Citation neighborhood

Papers in the corpus that this work cites (lower rings, blue) and that cite this one (upper rings, green). Dot size scales with the paper's in-corpus citation count — bigger dot = more influential within the endo/adeno field. Click a dot to open that paper. [ expand to 2 hops ] — adds papers reached through this work's immediate citers/citees. Heavier; up to 60 extra dots.

References (27)

Source provenance

openalex
last seen: 2026-11-20T06:09:37.170340+00:00
License: CC0 · commercial use OK