This retrospective study was conducted at the Institute of Neurology, Anhui University of Chinese Medicine, from January 2020 to July 2025. A total of 114 pregnant women with a confirmed diagnosis of WD (Leipzig score ≥4) were included. To ensure comparability and balance of data across different groups, patients were matched at a 1:1 ratio, resulting in 57 patients in the APO group and 57 patients in the UP group. The diagnosis of WD was independently established by two senior neurologists based on the Leipzig diagnostic criteria (2, 19), in accordance with a comprehensive assessment of clinical symptoms, laboratory tests, and genetic testing. This study was supported by the Clinical Hospital Research Ethics Committee of the Anhui University of Chinese Medicine (2024-SYSFYSY-17).

The inclusion criteria were as follows: (1) Pregnant women with a confirmed diagnosis of WD. (2) First pregnancy. (3) Provision of informed consent. (4) Aged ≥ 18 years old.

The exclusion criteria included: (1) Presence of other chronic liver diseases (e.g., viral hepatitis, autoimmune hepatitis). (2) History of severe renal insufficiency. (3) Not a first pregnancy. (4) Patients subjected to gynecological treatment. (5) Inability to comply with the study protocol.

Patients were divided into two groups according to strict criteria:

The UP group (n = 57): Patients delivered a live infant at ≥37 weeks of gestation, with neonatal birth weight between the 10th and 90th percentiles, and without hypertensive disorders of pregnancy, gestational diabetes mellitus, or clinical placental abruption.

APO group (n = 57): The composite outcome included the occurrence of one or more of the following: spontaneous abortion, missed abortion, stillbirth, and fetal developmental abnormalities.

Demographic and clinical parameters were recorded, including age at first pregnancy, gender, disease duration, disease subtype, pregnancy outcomes, and whether standardized copper-elimination therapy had been administered prior to pregnancy. Additionally, Kayser–Fleischer rings, ultrasound detection of hepatic fibrosis, and magnetic resonance imaging (MRI) detection of cranial abnormalities were routinely assessed.

Microelements, including serum copper, ceruloplasmin, urine copper (the last recorded data before pregnancy), serum zinc, and serum iron, were measured. Serum biochemical indices, liver function parameters, and hepatic fibrosis markers, including white blood cell (WBC) count, red blood cell (RBC) count, platelet, alanine aminotransferase (ALT), aspartate aminotransferase (AST), γ-glutamyl transpeptidase (γ-GT), total protein, total bilirubin, hyaluronic acid, laminin, procollagen III, and collagen IV, were detected using standard automated biochemical analyzers (Hitachi) in the institution’s central laboratory.

Seventy percent of the dataset was randomly allocated for training ML models, 15% for testing the models, and the remaining 15% for validation set. For WD patients without hepatic fibrosis, the sample size was extremely limited (n = 22). Therefore, 5-fold cross-validation was employed to evaluate the performance of ML models, and this approach was used as a sensitivity analysis. Four ML models, including deep learning (DL, fully connected deep neural network was used in this study), random forest (RF), gradient boosting machine (GBM), and generalized linear model (GLM), were utilized to distinguish between UP and APO in patients with WD. Detailed descriptions of ML model can be found in Supplementary materials, as previously published article (16).

Given the potential for multicollinearity among the independent variables, feature selection was performed prior to constructing ML models. Firstly, univariate analysis was conducted for all independent variables. Since variables were entered into the model individually at this stage, multicollinearity was not a concern. Secondly, variables with significant differences in the univariate analysis were further analyzed using least absolute shrinkage and selection operator (LASSO) regression. LASSO is widely used for screening variable due to its ability to handle multicollinearity. Subsequently, variables identified as significant by LASSO regression were further analyzed using multivariate analysis to determine independent risk factors associated with APOs. Finally, based on the independent risk factors, four ML models were constructed for the diagnosis and prediction of APOs.

Continuous variables were assessed for normal distribution using the Kolmogorov–Smirnov test. Data comparisons between the UP group and the APO group were conducted using the independent T-test (for normal distribution data) and the Mann–Whitney U test (for skewed distribution data), respectively. Categorical variables were described using sample sizes and percentages, and comparisons were analyzed using the chi-square test. Variables demonstrating significant difference (i.e., p < 0.05) in the univariate analysis were included in LASSO regression to identify independent risk factors for APOs. Subsequently, the prediction outcomes were visualized using a nomogram. Furthermore, four previously described ML models (16) were developed based on the identified significant variables. The predictive performance of all models was assessed using precision-recall (PR) curves, receiver operating characteristic (ROC) curves, and their respective area under the curve (AUC) values (i.e., AUCPR and AUCROC). As previously described (16), the changes of mean square error (MSE), root mean square error (RMSE), R square, as well as brier score, calibration-in-the-large, and calibration slope were calculated to evaluate model fitting performance. Stratified analyses were conducted for patients with hepatic fibrosis (with vs. without fibrosis) and cerebral function (abnormal vs. normal). All statistical analyses were performed using R software R software (XGBoost package, version 4.4.0), with p < 0.05 considered statistically significant.

Overall, a total of 114 patients with WD were included, comprising 57 WD patients with the UP and 57 WD patients with APO. Of 57 patients with APO, 32 cases involved fetal death, 18 cases involved spontaneous abortion, 5 cases involved fetal malformation, and 2 cases involved stillbirth (Figure 1; Table 1). No significant difference was observed in age at first pregnancy between UP and APO groups, but the disease duration in the APO group was shorter than that in the UP group (Z = −2.242, p < 0.001, Table 1). The APO group was predominantly observed in patients with the cerebral function of WD, whereas the UP group was mainly occurred in the hepatic fibrosis (χ² = 23.392, p < 0.001, Table 1). Additionally, the APO group had not undergone adequate decoppering therapy prior to pregnancy (χ² = 52.627, p < 0.001, Table 1) and exhibited pronounced clinical manifestations, including Kayser-Fleischer rings (χ² = 28.977, p < 0.001, Table 1), hepatic fibrosis (χ² = 5.632, p = 0.031, Table 1), and cerebral function (χ² = 7.983, p = 0.008, Table 1).

In this study, the levels of microelements, including serum copper, ceruloplasmin, 24-h urinary copper, serum iron, and serum zinc, were measured (Table 2). The results showed that the levels of 24-h urinary copper (Z = −4.123, p < 0.001, Table 2), serum iron (Z = −2.596, p = 0.006, Table 2), platelet (Z = −3.160, p = 0.002, Table 2), ALT (Z = −2.693, p = 0.007, Table 2), AST (Z = −2.901, p = 0.004, Table 2), GGT (Z = −2.098, p = 0.038, Table 2), and total bilirubin (Z = −5.733, p < 0.001, Table 2) were elevated, whereas WBC (t = 2.038, p = 0.044, Table 2) and total protein (Z = 5.528, p < 0.001, Table 2) were decreased in the APO group compared to the UP group. Furthermore, hepatic fibrosis biomarkers, excepting for laminin (Z = −0.955, p = 0.340, Table 2), were significantly elevated in the APO group than in the UP group.

A LASSO regression analysis was performed on the aforementioned 16 significant variables. The final model left 11 variables, including hepatic fibrosis, cerebral function, pre-pregnancy copper-chelating therapy, WBC, platelet, total bilirubin, total protein, AST, hyaluronic acid, procollagen III, collagen IV, 24-h urinary copper, and serum iron (Figures 2A,B). These results were also visualized using a nomogram (Figure 2C). Among these variables, total protein demonstrated the highest predictive value (AUC = 0.823, Table 3; Figure 2D), while WBC showed the lowest value (AUC = 0.637, Table 3; Figure 2E).

The performance of the four ML models is summarized in Tables 4, 5 and Figures 3A–H. The GLM achieved the highest accuracy (0.850) and demonstrated optimal model performance in AUCPR (0.998) and AUC (0.987) (Table 4; Figures 3G,H), suggesting that its predictions are the most reliable and accurate. Minimal differences were observed for all metrics (e.g., accuracy, AUC, and sensitivity) between the testing set, training set, and validation set of the GLM model (Table 4), which reflects extremely high stability and an absence of overfitting (Table 5). The RF exhibited an accuracy (0.800) comparable to that of DL and also had the highest AUCPR (0.998) and AUC (0.987) (Table 4; Figures 3C,D). The GBM had perfect accuracy (0.986), AUC (1.000) in the training set (Tables 4, 5; Figures 3E,F). but declined in the testing set (accuracy: 0.750, Tables 4, 5), suggesting that the model over-memorized the training data, leading to poor performance on new data. The DL model showed a high AUC (1.000) in the testing set but a negative R square (Tables 4, 5; Figures 3A,B), suggesting that poor generalization ability and unreliable results of the DL model. The performance evaluations of ML models are shown in Table 5 and Supplementary Table 1.

Of the 114 patients with WD, 51 had normal cerebral function and 63 had abnormal cerebral function. Based on the findings of univariate analysis (Supplementary Tables 2, 3), LASSO regression and ROC analysis identified disease duration, pre-pregnancy treatment, total bilirubin, total protein, collagen IV, Kayser-Fleischer rings, and serum iron as independent risk factors in WD patients with normal cerebral function (Figures 4A,B). The GLM achieved the best performance on the testing set, including accuracy (1.000), AUC (1.000), and sensitivity (1.000) (Figures 4I,J; Supplementary Table 4), indicating highly reliable classification capability. All metrics demonstrated high consistency across the training, validation, and testing sets, with no significant overfitting (Supplementary Tables 5, 6).

The DL model ranked second. It achieved perfect classification metrics (accuracy, AUC, and sensitivity all 1.000) in the testing set (Supplementary Table 4; Figures 4C,D), indicating strong classification capability, along with high accuracy in regression prediction (Supplementary Tables 4–6). However, some metrics (e.g., accuracy 0.939) were slightly lower in the training set than those in the testing set (Supplementary Tables 4, 5). The RF model had substantially lower accuracy (0.800) and sensitivity (0.500) on the testing set compared to other models (Supplementary Table 4; Figures 4E,F), indicating limited classification ability for new samples. The GBM showed significantly higher regression error in the testing set (Supplementary Tables 5, 6; Figures 4G,H), suggesting overfitting.

In WD patients with abnormal cerebral function, LASSO regression and ROC analysis were conducted on variables with significant differences in the univariate analysis (Supplementary Tables 2, 3; Figures 5A,B). The GBM model achieved the highest accuracy (1.000), AUC (1.000), and sensitivity (1.000) in the testing set (Supplementary Tables 4–6; Figures 5G,H), indicating its superior and stable performance. The GLM model demonstrated favorable accuracy (0.833), AUC (0.889), and AUCPR (0.958) in the testing set (Supplementary Tables 4–6; Figures 5I,J), with stable regression metrics. Whereas DL and RF models showed significantly lower performance on key metrics such as accuracy (0.643) and sensitivity (0.482) in the training set compared to the testing and validation sets (Supplementary Tables 4–6; Figures 5C–F), suggesting the presence of overfitting in these models.

Of 114 patients with WD, 92 cases with and 22 without hepatic fibrosis. For WD patients with hepatic fibrosis, LASSO regression and ROC analysis were conducted on the variables with significant differences in the univariate analysis (Supplementary Tables 2, 3; Figures 6A,B). Both DL and GLM achieved the best accuracy (0.857), AUC (0.911), and AUCPR (0.974), with all metrics demonstrating high consistency across the training, testing, and validation sets (Supplementary Tables 4–6; Figures 6I–J). These findings suggested that the two models have superior, reliable and stable performance on new data. The GBM had moderate accuracy (0.857) and AUCPR (0.944) in the testing set but exhibited overfitting (Supplementary Tables 4–6; Figures 6A,B). The RF achieved perfect scores for all metrics in the training set (accuracy: 1.000, AUC: 1.000), but these metrics declined significantly in the testing set (accuracy: 0.625, AUC: 0.823), accompanied by a negative R square value (Supplementary Tables 4–6), indicating severe overfitting and poor generalization capability.

Although this study only enrolled 22 WD patients without hepatic fibrosis, we proceeded to develop ML models. Cerebral function and Kayser-Fleischer rings demonstrated statistical significance in the univariate analysis, followed by LASSO regression, ROC analysis, and ML models (Supplementary Tables 1–6; Figures 7A–J). However, all models performed bad performance in the testing set and exhibited significant overfitting. In addition, sensitivity analysis based on 5-fold cross-validation identified the GLM as the optimal model, achieving the best classification performance and predictive calibration, with the lowest risk of overfitting. In contrast, RF and GBM demonstrated significant overfitting (Supplementary Table 7).