This was a single-center, retrospective cohort study of consecutive women who presented with EMs and a comparative group with AD who attended Shuguang Hospital Affiliated with Shanghai University of Traditional Chinese Medicine (TCM). The diagnostic accuracy of the EMs and AD groups was evaluated based on a retrospective study design.

Patients with EMs or AD were enrolled in the study through the Shuguang Hospital Affiliated with Shanghai University of TCM electronic clinical database. The recruitment process followed several steps to identify and select suitable candidates with complete and relevant data. (1)

Patient records: Potential participants with EMs or AD were identified by reviewing the medical records of patients who attended the obstetrics and gynecology inpatient departments at Shuguang Hospital affiliated with Shanghai University of TCM, between January 2016 and December 2023.

This study investigated the potential value of a comprehensive set of key hematological and biochemical parameters in the diagnosis and prediction of EMs and AD. Venous blood samples were collected from patients and analyzed using an automated hematology analyzer for complete blood cell counts, including red blood cell count, white blood cell count, and platelet count. Additionally, platelet indices and coagulation function markers were measured using chemiluminescence immunoassay techniques.

In the statistical analysis that was conducted using SPSS version 26.0, the significance threshold was set at α = 0.05. Because the data in this study did not adhere to a normal distribution, the results are presented as medians with the 25th and 75th percentiles [M (Q25, Q75)], and the Mann‒Whitney U test was applied for intergroup comparisons. The Delong test was used to assess differences in the area under the curve (AUC) between the models. A P-value less than 0.05 (P < 0.05) was interpreted as indicating a significant difference between the groups under analysis.

Feature selection helps remove irrelevant features to prevent overfitting. Feature selection was conducted before ML modeling to reduce data dimensionality, enhance model training efficiency and predictive performance, and improve the generalization ability of the model to new data (26). Within the EMs and AD groups, data demonstrating significant discrepancies underwent normalization via the Z score technique to mitigate variances across numerical scales. Three ML algorithms were used to screen for hematological feature factors: logistic regression (LR), support vector machine (SVM) classification, and K-nearest neighbors (KNN) classification. The specific parameters are detailed in Supplementary Table S1. LR classification measured the importance of feature variables through the coefficients obtained after L1 regularization. SVM classification and KNN classification both assessed the importance of features using the Recursive Feature Elimination (RFE) method, where feature importance was determined by cumulative weight values. Subsequently, the feature factors selected through machine learning were cross-validated to identify the optimal feature factors for use in further research. A Venn diagram was generated using jvenn (27).

In this study, we conducted analyses using five classic ML models: the LR classifier, eXtreme Gradient Boosting (XGBoost) classifier, multilayer perceptron (MLP) classifier, SVM classifier, and random forest (RF) classifier. Our methodology divided the dataset into training (80%) and validation (20%) sets. This approach, which is designed to sequentially rotate the test set, enhances the reliability of our results by reducing random variance. For each algorithm, we adopted a rigorous training regimen using fivefold nested cross-validation on the training dataset. The determination of the optimal parameters for each model was facilitated through a comprehensive grid search, the specifics of which are listed in Supplementary Table S2. Building on this foundation, we enhanced our methodology by developing a stacked ensemble model. The model, which integrated three selected basic classifiers, was constructed with LR as the stacking algorithm of the meta-classifier. This integration aimed to enhance the accuracy and generalizability of our results.

The LR classifier is recognized as a classic and commonly utilized model for risk prediction because of its simplicity in model configuration, rapid training speed, and excellent interpretability (28). The XGBoost classifier has emerged as a popular ML algorithm that is renowned for its high performance and flexibility, serving as an effective implementation of the gradient boosting framework (29). The MLP classifier, which is also a classic in the ML domain, employs backpropagation to train the network, calculates the error between the actual and predicted outputs and adjusts the weights by propagating this error back through the system. Thus, it is highly effective for complex problem solving (30). The SVM classifier, which leverages kernel functions, achieves linear separation in high-dimensional space and is recognized for its stability (31). The RF classifier is known for its robustness. It operates by constructing numerous decision trees during training and deriving the class by the mode of the classes of individual trees for classification tasks. This model has demonstrated its effectiveness across a variety of classification problems (32). Stacked ensemble algorithms is a widely applied ensemble learning method that combines basic classifier models to yield predictions with higher accuracy and better generalization capabilities (33).

Model performance was assessed through receiver operating characteristic (ROC) analysis, and the area under the curve (AUC) and 95% confidence interval (CI) were used as the key metrics for evaluating model efficacy. Then, the accuracy, sensitivity, and specificity were calculated. In addition, a calibration curve was used to evaluate the model performance.

All ML processes were carried out using Python 3.7 with several essential libraries: Scikit-learn (1.1.3) for implementing machine learning models, including Logistic Regression with L1 regularization, SVM, and KNN. Pandas (1.2.4) for data manipulation. NumPy (1.20.2) for numerical computations.

The study cohort included a total of 558 patients, 329 in the EMs group and 229 in the AD group. The age range of the participants in this study ranged from 22 to 67 years. Specifically, the age in the EMs group was 33 (29, 39) years, and in the AD group, it was 34 (30, 39) years. There was no significant difference in age distribution between the two groups (P > 0.05), indicating that the hematological indices were comparable. The hematological indices of the participants, which included complete blood cell counts, platelet indices, and markers of coagulation function, are presented in Table 1. Analysis of these indices revealed significant differences in several parameters (P < 0.05), including white blood cells (WBCs), mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCHC), red blood cells (RBCs), platelets (PLTs), red cell distribution width (RDW), mean corpuscular hemoglobin (MCH), basophilic granulocytes, plateletcrit (PCT), monocytes, international normalized ratio (INR), activated partial thromboplastin time (APTT), prothrombin time (PT), fibrin degradation product (FDP), and antithrombin III (AT-III), between the EMs and AD groups.

This study selected differential hematological indices from cohort characteristics for feature factor screening. The LR algorithm of L1 regularization was used to filter important variables. The top ten variables identified were RDW, APTT, PCT, MCHC, PT, INR, MCV, PLTs, RBCs, and AT-III. For further analysis, a SVM classification algorithm was employed, revealing the top ten variables: RDW, INR, APTT, PT, MCH, RBCs, MCHC, monocytes, FDP, and AT-III. Additionally, the KNN classification algorithm was utilized to identify the top ten variables: RDW, INR, basophilic granulocyte, AT-III, APTT, WBCs, MCHC, monocytes, MCH, and PLTs. The importance coefficients of the feature factors identified by these three feature selection methods are illustrated in Figures 1–3. A Venn diagram was constructed to identify the intersection of feature factors derived from the LR, SVM, and KNN algorithms. The intersection targets for the three datasets were RDW, APTT, MCHC, INR, and AT-III (Figure 4).

First, five classic ML models were developed and validated, and the model performances are listed in Tables 2, 3 and Figures 5–7. Among these, the model with the best performance on the training set was the XGBoost classification model (AUC) = 0.865, 95% CI: 0.832–0.899), followed by the RF classification model (AUC = 0.816, 95% CI: 0.776–0.855), the MLP classification model (AUC = 0.731, 95% CI: 0.683–0.778), the LR classification model (AUC = 0.725, 95% CI: 0.678–0.773), and the SVM classification model (AUC = 0.724, 95% CI: 0.676–0.772). The XGBoost classification model also showed the best performance in the validation set (AUC = 0.747, 95% CI: 0.652–0.842), followed by the MLP classification model (AUC = 0.744, 95% CI: 0.650–0.839), the LR classification model (AUC = 0.735, 95% CI: 0.640–0.831), the RF classification model (AUC = 0.731, 95% CI: 0.634–0.827), and the SVM classification model (AUC = 0.727, 95% CI: 0.631–0.824).

Subsequently, the ROC curves of the five ML methods were tested using the DeLong test. The results indicated that there was no statistically significant difference in the ROC curves of the aforementioned machine learning models (P > 0.05), as shown in Table 4. Calibration curves of the validation set for multiple models (Figure 8) demonstrated that the predicted probabilities of the five machine classification models were close to the actual probabilities.

A stacked ensemble model was constructed by selecting the three best-performing base learners (XGBoost, RF, and MLP). The stacked ensemble model was built on the second-layer LR meta-model based on the first layer of base learners, with the following model parameters: regularization factor: 1, number of iterations: 100, type of regularization: l2, convergence metric: 0.0001. Compared to XGBoost (AUC = 0.754, 95% CI: 0.647–0.860; Specificity = 0.669), RF (AUC = 0.778, 95% CI: 0.673–0.883; Specificity = 0.846), and MLP (AUC = 0.802, 95% CI: 0.698–0.906; Specificity = 0.863), the stacked ensemble model achieved an AUC = 0.803 (95% CI: 0.701–0.904; Specificity = 0.875), as detailed in Table 5 and Figure 9. These results showed that the ensemble model outperformed the individual models in terms of classification accuracy for EMs and AD, with improved performance and stronger generalizability.

EMs and AD are distinct but closely related conditions involving the presence of endometrial tissue outside the uterine lining. EMs is a chronic inflammatory disease that significantly impairs quality of life, often causing cyclic pain and infertility. AD is characterized by the invasion of endometrial tissue into the myometrium, leading to myometrial hypertrophy (34, 35). Currently, the definitive diagnosis of both conditions relies on invasive surgical or pathological examination, which is not always feasible. This underscores the need for non-invasive diagnostic methods to improve patient management (36). Recent studies have explored the use of ML for non-invasive diagnosis of EMs and AD. Guerriero et al. (37) used LR to differentiate EMs from AD based on ultrasound imaging. Balica et al. (38) employed five ML models (Xception, Inception-V4, ResNet50, DenseNet, and EfficientNetB2) to assist in ultrasound diagnosis, achieving an AUC of 90% and an accuracy of 80%. Raimondo et al. (39) developed a deep learning model for ultrasound-based diagnosis of AD, noting its potential to reduce overdiagnosis. However, ultrasound diagnosis is heavily dependent on the examiner's expertise, which can result in missed or incorrect diagnoses, particularly in early or deep pelvic lesions. Our study offered a novel approach by integrating readily accessible peripheral blood and coagulation markers into a stacked ensemble ML model. This approach enhanced the accuracy and reliability of differentiating between EMs and AD. It addressed a critical gap in the non-invasive diagnosis of these conditions and contributes to early diagnosis and personalized treatment strategies.

In this study, we demonstrated the feasibility and sensitivity of applying ML methods to screen for characteristic factors, including complete blood counts, platelet indices, and coagulation markers, that could differentiate between EMs and AD. Five features were selected as hematological indicators for assessing EMs and AD and were identified as key potential factors for discriminating between the two diseases. Furthermore, based on five classic ML models, we identified the three with the best performance for the construction of a stacked ensemble model. The stacked model emerged as the optimal model for distinguishing between EMs and AD (AUC = 0.803, 95% CI: 0.701–0.904).

Considering the computational resources wasted on redundant and irrelevant features within the original feature set during model training and prediction (40), we employed three ML methods (LR, SVM, and KNN) for feature selection and performed cross-validation to retain useful feature factors. L1-regularized LR improves model efficiency and generalization by reducing the number of features, as it retains only those features that significantly contribute to predictions while eliminating noise (41). In contrast, SVM and KNN excel at handling non-linear relationships. When using RFE, they can recursively eliminate the least impactful features, allowing the models to focus on the most relevant aspects of the data, thereby enabling faster and more accurate predictions (42). The combined use of the three methods effectively leverages the strengths of different models. Precise feature factor selection not only enhances model performance but also increases model transparency and interpretability (43). More intriguingly, this study identified key features, particularly RDW, MCHC, APTT, INR, and AT-III, through machine learning techniques, highlighting the hematological differences between patients with EMs and patients with AD. These findings underscore the potential differences in bleeding and coagulation between the two conditions.

Our study revealed that, as significant characteristic factors in the peripheral blood for these two diseases, the RDW in patients with AD was greater than that in patients with EMs, and the MCHC in patients with AD was lower than that in patients with EMs. RDW represents the standard deviation or coefficient of variation percentage of RBC volume, indicating significant size disparities in certain anemias. An increase in RDW reflects a severe disruption in erythrocyte homeostasis, including impaired erythrocyte production and abnormal erythrocyte survival (44). The MCHC is a critical indicator among the erythrocyte parameters that often suggests anemia when decreased (45). Dugdale et al. (46) reported a negative correlation between RDW and hemoglobin concentration over several months. An increase in RDW precedes a clinically significant decrease in hemoglobin levels by weeks; therefore, RDW is recommended as a valuable routine marker for the early detection of iron deficiency anemia. AD patients often exhibit clinical manifestations of increased menstrual flow (7), but EMs patients do not. This may explain the differences in the RDW and MCHC, suggesting a greater likelihood of bleeding and a predisposition to anemia in AD patients.

In our research, we identified three specific coagulation factors related to clotting: APTT, INR, and AT-III. The APTT and INR in the AD cohort were lower than those in the EMs cohort, and the AT-III levels were greater in the AD cohort. The APTT is a critical parameter for coagulation that is commonly used to predict bleeding tendencies and hypercoagulable states. A decrease in APTT is associated with hypercoagulability, indicated by an increase in thrombin generation and a greater risk of thrombosis (47). INR is a standardized PT that adjusts for variations in coagulation activator reagents, allowing the PT values measured by different laboratories and reagents to be comparable. A lower INR suggests an increased risk of thrombosis (48). This study demonstrated that patients with AD have a greater hypercoagulable state and greater thrombotic risk than patients with EMs. Lin et al. (49) reported that APTT is decreased in AD patients, indicating a hypercoagulable state. Yang et al. (24) reported a significant decrease in APTT among AD patients with thrombosis. A study indicated that shorter APTT in EMs patients might be related to a potential hypercoagulable state associated with the disease, and the role of the local coagulation system in the disease pathogenesis cannot be excluded (50). Lin et al. (49) also showed a negative correlation between coagulation markers and hemoglobin in AD patients with anemia. Anemia can affect coagulation parameters and increase the risk of thrombus formation (51). Yamanaka et al. (52) suggested that the coagulation dysfunction caused by AD could be a possible reason for thrombosis formation and menorrhagia. Menorrhagia can lead to anemia, further promoting a hypercoagulable state and possibly leading to thrombosis in a vicious cycle (49). Based on these studies, we hypothesize that menorrhagia-induced anemia and subsequent hypercoagulable changes are the reasons for the specific differences in coagulation markers between patients with AD and patients with EMs, but the related underlying physiological mechanisms require further investigation. Interestingly, AT-III levels were greater in the AD cohort than in the EMs cohort, which seems inconsistent with the trends in APTT and INR. EMs is an inflammatory disease characterized by increased expression of inflammatory and angiogenic factors (53). The peritoneal fluid of patients with EMs contains high levels of macrophages and immune cells, which secrete cytokines, angiogenic factors, and growth factors (54–56). AT-III is a nonvitamin K-dependent protease that regulates coagulation and inhibits inflammation within the endothelium (57, 58). Therefore, we speculate that AT-III is activated in the inflammatory response caused by the abnormal growth of endometrial tissue in EMs patients and plays an anti-inflammatory role by increasing its consumption within the vasculature.

Compared with traditional clinical diagnostic prediction models, the stacked ensemble model based on multiple machine learning algorithms successfully developed by us exemplifies a powerful and flexible strategy in machine learning. It integrates numerous learners and strategically adjusts using a meta-learner. Thus, this model achieves exceptional classification performance and significantly enhances the ability to generalize to new data (59). While stacked ensemble models are widely used in other areas, their application to the noninvasive diagnosis of EMs and AD has not been extensively explored. This study applies a stacked ensemble model in this specific clinical context, contributing to the understanding of its potential in this area. Our research demonstrates that the stacked ensemble model outperforms traditional models by achieving a higher AUC (0.803) and excelling in specificity (0.875). This consistent advantage underscores its potential for more accurate and reliable differentiation between EMs and AD. The model's balanced performance in both accuracy and specificity highlights its robustness and suitability for clinical applications where precision and reliability are crucial.

Furthermore, peripheral blood and coagulation markers are cost-effective and readily available indicators. Thus, our study was able to screen for early classification predictions of EMs and AD without the need for invasive diagnostic methods. This advancement aids clinicians in the early identification of high-risk patients and in taking timely measures to address relevant risk factors, thereby devising more rational diagnostic and therapeutic plans. It is important to note the limitations of our study. Our research was conducted as a single-center study, and further research will need to collect data from patients across different countries or regions to enhance the generalizability of the research findings. While it is effective in differentiating between EMs and AD, further development and validation would be required to extend its application to a broader diagnostic context. Expanding the model's capability to identify these conditions among a wider range of possible diagnoses could significantly advance computer-assisted diagnostic tools.