Figure Legends 
Figure 1. Flowchart of participant selection from the MIMIC-IV database (version 3.1, 2008–2022). Of 4,756 pregnant women identified from hospitalization records with a gestational age of ≥20 weeks, 4,247 were excluded because of missing values in any analysis variable (complete-case analysis), leaving 509 women for analysis: 78 preeclampsia cases (15.3%) and 431 controls (84.7%). Preeclampsia was defined by ICD-9-CM 6424x/6425x/6427x and ICD-10-CM O11x/O14x diagnosis codes (operational definition as described in the Methods). The analysis cohort was randomly split, stratified by outcome, into a training set (n = 355, 69.7%) and a validation set (n = 154, 30.3%).

Figure 2. Restricted cubic spline analysis of the association between the serum uric acid-to-creatinine ratio (UA CR) and preeclampsia (n = 509). Panels show (A) the crude model, (B) Model 1 adjusted for age, and (C) Model 2 adjusted for age, language, and red blood cell count. Solid lines denote adjusted odds ratios (ORs) and shaded bands the 95% confidence intervals, with the reference set at the cohort median UA CR of 7.83 (OR = 1 by construction); three knots were placed at the 10th, 50th, and 90th percentiles of the UA CR distribution. The association was positive and approximately linear in all models (P overall and P nonlinearity: A, 0.004 and 0.932; B, 0.01 and 0.92; C [Model 2], 0.001 and 0.911). ORs are plotted on the log scale.

Figure 3. Spearman correlation heatmap among the continuous features of the final LASSO-selected machine-learning feature set (UA CR, alanine aminotransferase, and hematocrit; n = 509). Marital status, the sole categorical feature, is not included in the correlation matrix. Cell labels give pairwise Spearman correlation coefficients (UA CR–ALT 0.11, UA CR–hematocrit 0.04, ALT–hematocrit 0.13); all pairwise |r| were < 0.15, indicating no material collinearity among the entered continuous features.

Figure 4. Discrimination, calibration, and clinical utility of the six machine-learning models (random forest, XGBoost, LightGBM, CatBoost, AdaBoost, and TabPFN; preeclampsia was the positive class throughout). (A) Receiver operating characteristic (ROC) curves on the training set (n = 355); (B) ROC curves on the validation set (n = 154), with area under the curve (AUC) values ranging from 0.655 (random forest) to 0.761 (LightGBM); (C) Platt-scaled calibration curves on the training set; (D) Platt-scaled calibration curves on the validation set, plotting predicted probabilities against observed event frequencies with the ideal line shown for reference; (E, F) model-metric dot plots (accuracy, balanced accuracy, sensitivity, specificity, precision, recall, F1, negative/positive predictive value, kappa, Matthews correlation coefficient, J index, detection prevalence, ROC AUC) on the training and validation sets, respectively; (G, H) decision-curve analysis showing net benefit versus threshold probability for each model on the training and validation sets, with "treat all" and "treat none" strategies as reference.

Figure 5. SHAP (SHapley Additive exPlanations) analysis of the training-set LightGBM model (n = 355). (A) Mean absolute SHAP values ranking the four LASSO-selected features (hematocrit 0.506, UA CR 0.482, alanine aminotransferase 0.429, marital status); (B) SHAP summary (beeswarm) plot showing the direction and distribution of each feature's contribution; (C) SHAP waterfall plot for a representative individual prediction; (D–G) SHAP dependence plots for UA CR, hematocrit, alanine aminotransferase, and marital status, respectively. SHAP values are descriptive rankings of contribution and do not establish incremental value of any single feature.

Figure 6. Forest plot of subgroup analyses for the association between UA CR (highest vs lowest quartile, Q4 vs Q1) and preeclampsia, estimated within strata of age (< 35 vs ≥ 35 years), marital status, race, and language (most-adjusted Model 2; overall-cohort quartile cutoffs retained within strata; the stratification variable itself was omitted as a covariate within its own strata). Squares denote stratum-specific ORs with 95% confidence intervals and pointwise P values; the overall estimate (OR 2.758, 95% CI 1.318–6.128) is shown for reference. The association was statistically significant only in the Hispanic stratum (OR 9.78, 95% CI 1.12–85.63; n = 49); the Asian stratum (n = 31) was not estimable owing to sparse events. P values for multiplicative interaction are shown at the bottom of each group of strata: race 0.026, age 0.441, marital status 0.697, language 0.632.

Figure S1. LASSO (least absolute shrinkage and selection operator) feature selection with 10-fold cross-validation on the training set (n = 355). (A) Selection frequency of each candidate variable across cross-validation folds (dashed line, high-frequency threshold); (B) heatmap of common variable-selection patterns (yellow, selected) with their occurrence counts. High-frequency variables correspond to the final four-variable feature set: UA CR, alanine aminotransferase, marital status, and hematocrit.