Enhancing the positive predictive value of early-stage ovarian cancer detection using a two-step machine learning framework

This study builds upon a previously reported dataset [ 15 ] by incorporating additional measurements of CA72-4 and applying alternative machine learning models. Briefly, serum samples were collected between 2014 and 2022 from patients with EOC, patients with benign ovarian tumors, and healthy controls at Tokai University Hospital (Kanagawa, Japan), Kyorin University Hospital (Tokyo, Japan), National Cancer Center Japan (Tokyo, Japan), St. Marianna University Hospital (Kanagawa, Japan), Kanagawa Cancer Center (Kanagawa, Japan), Tohoku Medical Megabank Organization (Miyagi, Japan), Tsukuba University Hospital (Ibaraki, Japan), Yokohama City University Hospital (Kanagawa, Japan), Fujita Health University Hospital (Aichi, Japan), LSI Medience Corporation (Tokyo, Japan), KAC (Kyoto, Japan), SOIKEN (Osaka, Japan), and Sunfco (Tokyo, Japan) (Table 1 ). Participants meeting any of the following criteria were excluded: (i) presence of multiple cancer types; (ii) history of hormonal therapy for malignant tumors, autoimmune diseases, or thyroid disorders; (iii) abnormal blood test results (white blood cell count > 9,600/µL, platelet count > 48 × 10⁴/µL, lactate dehydrogenase > 263 U/L, hemoglobin 3.0 mg/dL); (iv) renal dysfunction; or (v) liver dysfunction. All participants provided written informed consent, and the study was approved by the institutional review board (IRB) of Tokai University (initial IRB registration number: 09R-082; September 17, 2009). The study protocol, including the CSGSA framework, was subsequently reviewed and approved as an amended protocol by the Tokai University Clinical Research Review Committee (approval number: 15R-237; approved June 17, 2019). All research methods and procedures followed the ethical principles outlined in the Declaration of Helsinki and fully complied with applicable institutional, national, and international guidelines. This study was conducted as an observational study. Table 1 Demographic and clinical characteristics of the study population Variable Training set ( n  = 1,130) Hold-out test set ( n  = 567) p -value (Train vs. Test)* Age , mean ± SD EOC 373 (55.8 ± 12.0) 180 (56.3 ± 12.4) 0.65 Non-EOC 757 (54.4 ± 12.0) 387 (54.3 ± 11.7) 0.89 Non-EOC conditions , n (%) 0.90 Endometrioma (EM) 43 (5.7%) 23 (5.9%) Healthy women (HE) 608 (80.3%) 316 (81.7%) Uterine myoma (LE) 57 (7.5%) 26 (6.7%) Ovarian cyst (OCY) 49 (6.5%) 22 (5.7%) Total 757 (100.0%) 387 (100.0%) Histological subtype of EOC , n (%) 0.10 Serous (SE) 111 (29.8%) 68 (37.8%) Endometrioid (EN) 92 (24.7%) 49 (27.2%) Clear cell (CC) 130 (34.9%) 48 (26.7%) Mucinous (MU) 27 (7.2%) 13 (7.2%) Unclassified (UC) 13 (3.5%) 2 (1.1%) Total 373 (100.0%) 180 (100.0%) FIGO stage distribution (EOC only) , n (%) 0.46 Stage I 191 (51.2%) 81 (45.0%) Stage II 38 (10.2%) 24 (13.3%) Stage III 123 (33.0%) 66 (36.7%) Stage IV 21 (5.6%) 9 (5.0%) Total 373 (100.0%) 180 (100.0%) Abbreviations: EOC, epithelial ovarian cancer; SE, serous; EN, endometrioid; CC, clear cell; MU, mucinous; UC, unclassified; EM, endometrioma; HE, healthy women; LE, uterine myoma; OCY, ovarian cyst. * p-values between the training and hold-out test sets were calculated using Welch’s t-test for age and the chi-square test for categorical variables. No statistically significant differences were observed, indicating that the random split produced comparable groups Demographic and clinical characteristics of the study population Abbreviations: EOC, epithelial ovarian cancer; SE, serous; EN, endometrioid; CC, clear cell; MU, mucinous; UC, unclassified; EM, endometrioma; HE, healthy women; LE, uterine myoma; OCY, ovarian cyst. * p-values between the training and hold-out test sets were calculated using Welch’s t-test for age and the chi-square test for categorical variables. No statistically significant differences were observed, indicating that the random split produced comparable groups A total of 1,697 participants, including 553 cases of EOC and 1,144 non-EOC controls (healthy individuals and patients with benign ovarian diseases), were randomly assigned to a training set (373 EOC and 757 non-EOC) and an independent hold-out test set (180 EOC and 387 non-EOC) using a random number generator with a fixed seed. As summarized in Table  1 , no significant differences in demographic or clinical characteristics were observed between the training and hold-out test sets, indicating that the groups were comparable. The training set was used for model development and optimization, including feature selection and hyperparameter tuning, whereas the hold-out test set was reserved exclusively for final model evaluation. CA125 and HE4 concentrations were measured as previously described [ 15 ], and CA72-4 levels were measured using an electrochemiluminescence immunoassay at LSI Medience Corporation, a clinical testing laboratory in Japan. To ensure assay reproducibility, both internal and external quality control procedures were strictly followed, with intra-assay coefficients of variation (CVs) typically maintained below 5%. The measurement ranges and detection limits were consistent with the manufacturer’s specifications. The preparation of EGPs from serum samples and subsequent LC–MS analysis were performed as previously described, without modifications [ 15 ]. Briefly, proteins were subjected to trypsin digestion, followed by ultrafiltration-based enrichment of glycopeptides. The expression levels of individual EGPs were then determined using a liquid chromatography quadrupole time-of-flight mass spectrometer (HP1200 + 6520; Agilent Technologies, Palo Alto, CA, USA) equipped with a C18 column (Inertsil ODS-4; 2 μm, 100 Å, 100 mm × 1.5 mm ID; GL Science, Tokyo, Japan). Mass spectrometry data were processed and normalized as previously described [ 15 ]. For machine learning analysis, principal component analysis (PCA) was performed to reduce the dimensionality of the EGP-derived features. The number of principal components (PCs) was fixed at 100 based on preliminary analyses using the training data. These analyses indicated that classifier performance (cross-validated ROC-AUC) plateaued around this value while retaining 77.4% of the total variance, beyond which model complexity and computational cost increased without meaningful gains. Age and tumor marker values (CA125, HE4, and CA72-4) were also included as input features. Machine learning analyses were conducted using Python (3.10.16) with the following libraries: LightGBM (4.6.0), XGBoost (2.1.4), Scikit-learn (1.6.1), NumPy (2.1.3), and Pandas (2.2.3). A total of 1,130 samples were assigned to the training set, whereas 567 samples were reserved as an independent hold-out test set for final model evaluation. Although the hold-out test set was independent of the training set, both datasets were derived from the same multicenter cohort. Hyperparameter tuning (e.g., max_depth, n_estimators, and learning_rate) was performed using grid search with internal cross-validation applied exclusively to the training data. Statistical comparisons of model performance were conducted using Student’s t-test based on accuracy values obtained from each fold in a 10-fold cross-validation. The optimized model was subsequently evaluated in the hold-out test cohort consisting of 180 EOC and 387 non-EOC samples to assess generalization performance on previously unseen data. Samples from the hold-out test set were not used during the model training or hyperparameter optimization. Two-step classification models were constructed as follows: In the first step, a LightGBM model was developed using age and tumor marker levels (CA125, HE4, and CA72-4) to stratify patients into three categories: high-risk (EOC), low-risk (non-EOC), and indeterminate. In the second step, cases classified as indeterminate were further analyzed using an additional LightGBM model that incorporated PCA-derived EGP features, together with tumor markers and age information, to reclassify the cases as EOC or non-EOC (Additional File Table S1 ). Model performance was assessed on the independent hold-out test set using accuracy, ROC-AUC, sensitivity, specificity, PPV, and negative predictive value (NPV). Prevalence-adjusted PPV and NPV were calculated using Bayes’ theorem based on predefined assumptions of disease prevalence in a population-based screening setting. The prevalence of EOC was assumed to be 30 cases per 100,000 women, based on statistics reported by the Japanese government in 2025. In addition, the prevalence of major benign gynecological conditions that may contribute to false-positive results was incorporated into the calculation, including endometrioma (EM: 1,000 per 100,000), uterine myoma (LE: 2,000 per 100,000), and ovarian cysts (OCY: 1,000 per 100,000) [ 15 ]. The remaining population was assumed to consist of healthy individuals. Expected counts were calculated as follows: TP = sensitivity × prevalence_EOC, FN = (1 − sensitivity) × prevalence_EOC, FP = Σ (false-positive rate_i × prevalence_i) across each benign condition and healthy individuals, TN = Σ ((1 − false-positive rate_i) × prevalence_i) across each benign condition and healthy individuals. PPV and NPV were then calculated as PPV = TP / (TP + FP) and NPV = TN / (TN + FN). All prevalence-adjusted values therefore represent theoretical projections under assumed population conditions and are not direct observations from the study cohort. In addition, clinically relevant operating points were evaluated by assessing sensitivity at predefined high specificity thresholds. Statistical analyses were performed using Python (version 3.10.16) unless otherwise specified. Receiver operating characteristic (ROC) curve analyses and area under the curve (AUC) calculations were conducted using the roc_auc_score function in Scikit-learn. Statistical significance was set at p  < 0.05.

We evaluated the diagnostic performance of the three tumor markers from three perspectives: (1) discrimination between EOC and non-EOC groups, including healthy individuals and patients with LE, OCY, or EM; (2) detection of early-stage EOC; and (3) discrimination of the four major histological subtypes of EOC (clear cell, mucinous, endometrioid, and serous carcinoma) from benign conditions (Table  1 ). ROC-AUC analysis revealed clear differences in diagnostic performance among the three markers. Following a comparison between patients with EOC and all benign controls, CA125 showed the highest ROC-AUC (0.899), followed by HE4 (0.814) and CA72-4 (0.731). In contrast, in the comparison between EOC and EM, CA125 showed lower performance (ROC-AUC: 0.667), whereas HE4 (0.743) and CA72-4 (0.752) showed higher values. This pattern is consistent with the known elevation of CA125 in endometriosis, which may reduce its specificity in this context. Although CA72-4 showed lower overall performance, it achieved the highest ROC-AUC for distinguishing EOC from EM, suggesting potential complementary value (Fig.  1 A). Stage-wise analysis showed that diagnostic performance increased with disease stage for all three markers. In Stage I disease, CA125 showed a better performance (ROC-AUC: 0.838) than HE4 (0.716) and CA72-4 (0.672), indicating its relative advantage in early-stage detection (Fig.  1 B). Fig. 1 Characterization of tumor marker profiles (CA125, HE4, and CA72-4). (A) Tumor marker distribution across EOC and non-EOC groups. Left: Histograms show the distribution of each marker (vertical: marker level, horizontal: number of individuals). Right: ROC curves and corresponding AUC values for distinguishing EOC from each control group. Control subgroups include healthy individuals, LE (uterine myoma), OCY (ovarian cysts), and EM (endometrioma). (B) Tumor marker responses by EOC stage. Left: Histograms for each tumor marker stratified by clinical stage (Stages I–IV). Right: ROC analysis comparing each stage with non-EOC cases. (C) Tumor marker responses across histological subtypes of EOC. Left: Histograms for each marker within subtypes. Right: ROC analysis comparing each subtype with non-EOC cases. Subtypes include CC (clear cell carcinoma), MU (mucinous carcinoma), EN (endometrioid carcinoma), and SE (serous carcinoma). (D–F) Scatter plots showing pairwise correlations between markers on a log 10 scale. (D) CA125 vs. HE4, (E) HE4 vs. CA72-4, (F) CA125 vs. CA72-4 Characterization of tumor marker profiles (CA125, HE4, and CA72-4). (A) Tumor marker distribution across EOC and non-EOC groups. Left: Histograms show the distribution of each marker (vertical: marker level, horizontal: number of individuals). Right: ROC curves and corresponding AUC values for distinguishing EOC from each control group. Control subgroups include healthy individuals, LE (uterine myoma), OCY (ovarian cysts), and EM (endometrioma). (B) Tumor marker responses by EOC stage. Left: Histograms for each tumor marker stratified by clinical stage (Stages I–IV). Right: ROC analysis comparing each stage with non-EOC cases. (C) Tumor marker responses across histological subtypes of EOC. Left: Histograms for each marker within subtypes. Right: ROC analysis comparing each subtype with non-EOC cases. Subtypes include CC (clear cell carcinoma), MU (mucinous carcinoma), EN (endometrioid carcinoma), and SE (serous carcinoma). (D–F) Scatter plots showing pairwise correlations between markers on a log 10 scale. (D) CA125 vs. HE4, (E) HE4 vs. CA72-4, (F) CA125 vs. CA72-4 Fig. 2 Construction and evaluation of machine learning models for ovarian cancer risk classification. (A) Model 1: A single-step classification model using only age, CA125, HE4, and CA72-4 as input features. Patients are classified into high-risk or low-risk groups in one stage. (B) Model 2: A single-step model using age, CA125, HE4, CA72-4, and 1,712 EGP features. The EGPs are dimensionally reduced to 100 components using principal component analysis (PCA) prior to modeling. (C) Model 3: A two-step classification framework. In the first step, patients are classified into high-risk, low-risk, or indeterminate groups using only age, CA125, HE4, and CA72-4. In the second step, only the indeterminate group is re-evaluated using the same clinical markers plus 100 PCA components derived from 1,712 EGPs. (D) Classification accuracy of the training dataset. Three models (Model 1–3) were tested using neural network (NN), XGBoost, and LightGBM algorithms. Bar plots indicate accuracy for each configuration. (E) SHAP summary plot illustrating feature contributions in the CSGSA EOC model for discrimination between EOC and non-EOC controls. The plot displays the impact of each feature on model output across all samples, ranked by mean absolute SHAP value. Each point represents an individual sample, with color indicating the feature value (red, high; blue, low). Positive SHAP values indicate increased contribution toward EOC classification, whereas negative values indicate contribution toward non-EOC classification. (F) Feature importance based on LightGBM gain values. Left: Gain contribution from the first step of the two-step model. Right: Gain contribution from the second step. (G) Effect of CA72-4 on ROC-AUC. Comparison of ROC-AUC values for models with (+) and without (−) inclusion of CA72-4. (H) Effect of CA72-4 on PPV. PPV values were calculated for models with (+) and without (−) CA72-4, assuming a sensitivity over 60% Construction and evaluation of machine learning models for ovarian cancer risk classification. (A) Model 1: A single-step classification model using only age, CA125, HE4, and CA72-4 as input features. Patients are classified into high-risk or low-risk groups in one stage. (B) Model 2: A single-step model using age, CA125, HE4, CA72-4, and 1,712 EGP features. The EGPs are dimensionally reduced to 100 components using principal component analysis (PCA) prior to modeling. (C) Model 3: A two-step classification framework. In the first step, patients are classified into high-risk, low-risk, or indeterminate groups using only age, CA125, HE4, and CA72-4. In the second step, only the indeterminate group is re-evaluated using the same clinical markers plus 100 PCA components derived from 1,712 EGPs. (D) Classification accuracy of the training dataset. Three models (Model 1–3) were tested using neural network (NN), XGBoost, and LightGBM algorithms. Bar plots indicate accuracy for each configuration. (E) SHAP summary plot illustrating feature contributions in the CSGSA EOC model for discrimination between EOC and non-EOC controls. The plot displays the impact of each feature on model output across all samples, ranked by mean absolute SHAP value. Each point represents an individual sample, with color indicating the feature value (red, high; blue, low). Positive SHAP values indicate increased contribution toward EOC classification, whereas negative values indicate contribution toward non-EOC classification. (F) Feature importance based on LightGBM gain values. Left: Gain contribution from the first step of the two-step model. Right: Gain contribution from the second step. (G) Effect of CA72-4 on ROC-AUC. Comparison of ROC-AUC values for models with (+) and without (−) inclusion of CA72-4. (H) Effect of CA72-4 on PPV. PPV values were calculated for models with (+) and without (−) CA72-4, assuming a sensitivity over 60% When diagnostic performance was examined across histological subtypes, all three markers showed the highest performance for serous carcinoma, possibly reflecting the higher proportion of advanced-stage cases in this group. For clear cell carcinoma, CA125 showed higher performance (ROC-AUC: 0.859) than HE4 (0.702) and CA72-4 (0.664). In contrast, for mucinous carcinoma, CA72-4 showed the highest ROC-AUC (0.811), whereas CA125 and HE4 showed lower values, indicating potential complementary value of CA72-4 in this subtype (Fig.  1 C). The 95% confidence intervals (CIs) for all ROC-AUC analyses are provided in Additional File Table S2 . We further examined the relationships among the three tumor markers. A strong positive correlation was observed between CA125 and HE4 (Fig.  1 D), whereas no clear correlation was detected between HE4 and CA72-4 (Fig.  1 E). The relationship between CA125 and CA72-4 varied according to CA125 levels. When CA125 values were below 1,000 IU/L (log₁₀ = 3), little correlation was observed; above this threshold, CA72-4 levels increased markedly with CA125. These findings indicate that CA72-4 may provide complementary information, particularly at higher CA125 levels (Fig.  1 F). Three machine learning models were evaluated in this study (Fig.  2 A–C). Model 1 incorporated three tumor markers (CA125, HE4, and CA72-4) and patient age. Model 2 included the same features as Model 1 and additionally incorporated 100 PCs derived from 1,712 EGPs to prevent overfitting. Model 3 employed a two-step structure. In the first step, tumor markers and patient age were used to classify individuals into high-risk, low-risk, or indeterminate groups. In the second step, PCA-derived glycopeptide features were applied to further evaluate cases categorized as indeterminate (Additional File Table S1 ), mimicking real-world clinical workflows. Patient demographics and clinical backgrounds were well balanced, with no significant differences observed between the two cohorts (Table  1 ). To reduce overfitting and provide internal validation, the dataset was divided into a training set and an independent hold-out test set generated from the same multicenter cohort. Model development, including hyperparameter tuning and model comparison, was performed using the training dataset. Final model performance was evaluated on the independent hold-out test set. Three machine learning algorithms—XGBoost, LightGBM, and Neural networks—were evaluated. Neural networks were applied only to Models 1 and 2. XGBoost constructs decision trees using a level-wise growth strategy, whereas LightGBM adopts a faster leaf-wise strategy with histogram-based optimization, which is suitable for high-dimensional data [ 22 ]. The classification accuracies obtained from combinations of Models 1–3 and the three machine learning algorithms are shown in Fig.  2 D. The p -values were calculated using Student’s t-test based on accuracy values from a 10-fold cross-validation. For each model, hyperparameters were optimized using a Python-based grid search within the training set prior to model evaluation. Across all algorithms, classification accuracy increased from Model 1 to Model 2 and from Model 2 to Model 3. XGBoost yielded comparable accuracies for Models 2 and 3, whereas LightGBM showed a significant increase in accuracy in Model 3 compared with Model 2 ( p  < 0.01). In Model 3, the threshold defining the indeterminate range was set between 0.001 and 0.999, resulting in approximately 60–70% of cases being classified as indeterminate and subsequently evaluated in the second step. This threshold range was determined through a grid search using the training dataset. Among the approaches evaluated, the two-step LightGBM model achieved the highest accuracy. Threshold sensitivity analysis demonstrated that narrowing the indeterminate range reduced the proportion of patients undergoing second-stage glycopeptide evaluation to 64%, 36%, and 19% for threshold ranges of 0.001–0.999, 0.01–0.99, and 0.05–0.95, respectively. This was accompanied by reductions in AUC (0.962, 0.940, and 0.903), sensitivity (65.0%, 32.0%, and 35.7%), and projected PPV (18.7%, 8.0%, and 8.9%, Additional File Table S3 ). To assess robustness, we additionally evaluated different numbers of PCs. The corresponding AUC values for 30, 50, 100, 150, and 200 PCs were 0.950, 0.956, 0.962, 0.959, and 0.962, respectively (Additional File Table S4 ). To compare the proposed framework with a clinically implementable baseline model, an additional logistic regression model incorporating age, CA125, HE4, and CA72-4 was evaluated using the same independent hold-out test set. The logistic regression model achieved an AUC of 0.889 (95% CI, 0.851–0.921), sensitivity of 45.0%, specificity of 99.9%, and projected PPV of 10.9%, whereas the proposed two-step LightGBM model achieved an AUC of 0.962 (95% CI, 0.942–0.979), sensitivity of 65.0%, specificity of 99.9%, and projected PPV of 18.7% (Additional File Table S5 ). This model was designated the “CSGSA EOC model,” and its performance was evaluated in subsequent analyses. To improve the interpretability of the two-step model, SHapley Additive exPlanations (SHAP) analysis was performed for representative cases within the hold-out test cohort. Analysis of feature contributions using SHAP revealed that CA125, HE4, and PC2 made the largest contributions to the model predictions (Fig.  2 E). Waterfall plots for representative cases—a false positive (EM0002), true negative (EM0006), true positive (EOC0015), and false negative (EOC0003)—are shown in Additional File Fig. S1 . In the false-positive case (Additional File Fig. S1 A), elevated CA125 levels and specific PCs (PC6 and PC10) strongly contributed to an increased predicted risk despite the absence of EOC. In contrast, the true-negative case (Additional File Fig. S1 B) demonstrated dominant negative contributions from several PCA-derived components, resulting in substantially decreased logit scores. The true-positive case (Additional File Fig. S1 C) showed strong positive contributions from CA125 and HE4, together with multiple PCA components, leading to a high prediction probability. Conversely, in the false-negative case (Additional File Fig. S1 D), although CA125 contributed positively, multiple PCA components exerted negative effects that collectively reduced the final prediction score to below the classification threshold. These findings indicate that the two-step CSGSA model integrates tumor markers and glycopeptide-derived PCA features in a complementary manner, and that model outputs are influenced by complex interactions among multiple features rather than reliance on a single biomarker. The misclassified cases (false-positive and false-negative cases) further demonstrate how competing contributions among tumor markers and PCA-derived glycopeptide features can influence the final prediction, highlighting the role of second-step refinement in modulating ambiguous predictions. The feature importance based on the Gain metric from the LightGBM algorithm is shown in Fig.  2 F. In the first step, CA125 and HE4 were the most influential features, whereas CA72-4 and patient age contributed the least. In the second step, CA125 and HE4 again showed strong impact, with PC2 and PC8, derived from 1,712 EGPs, also making substantial contributions. The confusion matrix for the hold-out test cohort (Table  2 ) showed 117 true positives, 63 false negatives, 2 false positives, and 385 true negatives. The PPV was 98.3% (117/119; 95% CI, 94.1%–99.5%) and the NPV was 85.9% (385/448; 95% CI, 82.4%–88.9%), with CIs estimated using the Wilson score method (Table  3 ). When these results were projected to population-level screening conditions using Bayes’ theorem and published prevalence data, the prevalence-adjusted PPV and NPV were estimated to be 18.7% and 99.99%, respectively (Table  3 ). We evaluated the contribution of the tumor marker CA72-4. Two versions of the two-step (Model 3) classifier were constructed: one including CA72-4 and the other excluding it (using only CA125, HE4, age, and 1,712 EGPs). Both models were assessed using the same hold-out test set. ROC analysis showed comparable overall discrimination (AUC 0.962 vs. 0.957; Fig.  2 G, Additional File Table S6 A). Table 2 Confusion matrix of the two-step CSGSA model for epithelial ovarian cancer CSGSA Sum Positive Negative EOC 117 63 180 non-EOC EM 2 21 23 HE 0 316 316 LE 0 26 26 OCY 0 22 22 Sum 119 448 567 Confusion matrix of the two-step CSGSA model for epithelial ovarian cancer Table 3 Prevalence-adjusted PPV, NPV, sensitivity, and specificity of the two-step CSGSA model Metric Observed (%) 95% CI Prevalence-adjusted (%)* Sensitivity 65.0 57.7–71.5 – Specificity 99.5 98.1–99.9 – PPV 98.3 94.1–99.5 18.7 † NPV 85.9 82.4–88.9 99.99 † * Prevalence-adjusted values were calculated using Bayes’ theorem based on the reported real-world prevalence of epithelial ovarian cancer and relevant benign gynecological conditions (see Methods) † The adjusted PPV and NPV represent theoretical estimates under the assumed population prevalence and were not derived from the hold-out test dataset Prevalence-adjusted PPV, NPV, sensitivity, and specificity of the two-step CSGSA model * Prevalence-adjusted values were calculated using Bayes’ theorem based on the reported real-world prevalence of epithelial ovarian cancer and relevant benign gynecological conditions (see Methods) † The adjusted PPV and NPV represent theoretical estimates under the assumed population prevalence and were not derived from the hold-out test dataset However, under prevalence-adjusted conditions with sensitivity greater than 60%, the model including CA72-4 achieved a PPV of 18.7% (Table  3 ), whereas the model without CA72-4 yielded 4.4% (Additional File Table S6 B; Fig.  2 H). This difference is attributable to the impact of false-positive rates in low-prevalence settings, where even small increases in false positives among non-EOC groups substantially reduce the PPV after prevalence adjustment. The corresponding observed sensitivity, specificity, PPV, and NPV with 95% CIs are summarized in Additional File Table S6 B. Although the improvement in AUC was modest, inclusion of CA72-4 improved PPV under clinically relevant screening conditions. To assess the impact of prevalence assumptions on these estimates, an additional sensitivity analysis was performed using ovarian cancer prevalence values ranging from 10 to 100 cases per 100,000 women. The projected PPV varied from 7.4% to 42.8% across this range (Additional File Table S7 ). These findings confirm that the projected PPV is influenced by the assumed disease prevalence, as expected. In the two-step LightGBM model (CSGSA EOC model), PC2 and PC8 derived from the 1,712 EGPs contributed substantially to model performance. We therefore examined the associations between these PCs and the underlying EGP features. The PCA score plot (Fig. 3 A) showed clearer separation between the EOC and non-EOC groups along PC2 than along PC8. Statistical analysis using Student’s t-test showed a highly significant difference in PC2 values between the two groups ( p = 1.4 × 10⁻ 43 ), supporting its role in distinguishing EOC from non-EOC samples (Fig. 3 B). PCA loadings for PC2, sorted in ascending order to visualize the relative contributions of individual EGPs, are shown in Fig. 3 C. The seven glycopeptides with the largest positive and negative loadings were selected for further analysis. Among the strongest positive contributors were glycopeptides derived from alpha-1-acid glycoproteins with tetra-antennary glycans and from haptoglobins with highly fucosylated structures, consistent with our previous findings [ 15 ]. Fig. 3 Analysis of principal component 2 (PC2) features contributing to second-step classification. (A) PCA score plot of EGPs. Each dot represents a sample, plotted by PC2 and PC8 scores. Red: EOC; Blue: non-EOC. (B) Distribution of PC2 and PC8 scores by group. Left: PC2 scores; Right: PC8 scores. Statistical comparison between EOC and non-EOC groups was performed using a Student’s t-test. (C) PC2 loading plot. All 1,712 EGPs were ranked by PC2 loading values. The seven glycopeptides with the largest positive and negative PC2 loadings are shown in enlarged bar charts, representing the most influential features contributing to the classification model. Glycopeptides identified using MS/MS are highlighted in dark-blue. Proteins with multiple high-contribution glycopeptides include transferrin, α1-acid glycoprotein, and haptoglobin. (D) Volcano plot comparing EOC and non-EOC. Each dot represents a glycopeptide, with log 2 fold change on the x-axis and –log 10 p -value on the y-axis. The largest positive and negative loading glycopeptides from (C) are highlighted in orange, and those with confident MS/MS identification are highlighted in brown. (E) Glycan structures of identified glycopeptides. Representative glycopeptides from haptoglobin, α1-acid glycoprotein, and transferrin are shown with their amino acid sequences and glycan compositions. Structural annotations include N-acetylglucosamine, mannose, galactose, fucose, and N-acetylneuraminic acid Analysis of principal component 2 (PC2) features contributing to second-step classification. (A) PCA score plot of EGPs. Each dot represents a sample, plotted by PC2 and PC8 scores. Red: EOC; Blue: non-EOC. (B) Distribution of PC2 and PC8 scores by group. Left: PC2 scores; Right: PC8 scores. Statistical comparison between EOC and non-EOC groups was performed using a Student’s t-test. (C) PC2 loading plot. All 1,712 EGPs were ranked by PC2 loading values. The seven glycopeptides with the largest positive and negative PC2 loadings are shown in enlarged bar charts, representing the most influential features contributing to the classification model. Glycopeptides identified using MS/MS are highlighted in dark-blue. Proteins with multiple high-contribution glycopeptides include transferrin, α1-acid glycoprotein, and haptoglobin. (D) Volcano plot comparing EOC and non-EOC. Each dot represents a glycopeptide, with log 2 fold change on the x-axis and –log 10 p -value on the y-axis. The largest positive and negative loading glycopeptides from (C) are highlighted in orange, and those with confident MS/MS identification are highlighted in brown. (E) Glycan structures of identified glycopeptides. Representative glycopeptides from haptoglobin, α1-acid glycoprotein, and transferrin are shown with their amino acid sequences and glycan compositions. Structural annotations include N-acetylglucosamine, mannose, galactose, fucose, and N-acetylneuraminic acid In contrast, the strongest negative contributors included transferrin-derived glycopeptides with bi-antennary glycan structures. When mapped onto a volcano plot comparing EOC and non-EOC groups, most of these glycopeptides were located in the upper-right or upper-left quadrants, indicating substantial differences between the two groups (Fig.  3 D). The proposed glycopeptide structures are illustrated in Fig.  3 E. The performance of the two-step LightGBM model (CSGSA EOC model) was evaluated using an independent hold-out test set. The model showed strong discriminative performance across several clinically relevant comparisons (Fig.  4 A). The ROC-AUC for distinguishing EOC from non-EOC was 0.962, exceeding the performance of CA125 alone (AUC = 0.899; Fig.  1 A), and was particularly high (0.971) following a comparison between patients with EOC and healthy individuals. In contrast, discrimination between EOC and EM represented the most challenging comparison, with an ROC-AUC of 0.879, although the model maintained higher performance than that observed for individual tumor markers. The distribution of model scores across EOC stages was then examined, and ROC-AUC values were calculated against the non-EOC group. In Stage I disease, the model achieved an ROC-AUC of 0.942, compared with 0.838 for CA125 (Fig.  1 B), indicating improved discrimination for early-stage ovarian cancer (Fig.  4 B). The EOC group was further stratified into four major histological subtypes—clear cell, mucinous, endometrioid, and serous carcinoma—and each subtype was compared with the non-EOC group. The model demonstrated consistently high performance across all subtypes, with ROC-AUC values of 0.950 for clear cell carcinoma, 0.984 for mucinous carcinoma, 0.958 for endometrioid carcinoma, and 0.969 for serous carcinoma (Fig.  4 C). The comparative diagnostic performance of CA125, HE4, CA72-4, and the CSGSA EOC model using ROC-AUC values is shown in Fig.  4 D–F. When comparing the EOC and non-EOC groups—including EM, OCY, LE, and healthy individuals—the CSGSA EOC model showed higher discriminative performance than the individual tumor markers across all control groups (Fig.  4 D). In stage-stratified analyses, the CSGSA EOC model demonstrated performance comparable to that of CA125 in Stage IV disease but showed improved discrimination in Stage I cases (Fig.  4 E). Similarly, in analyses stratified by histological subtype, the CSGSA EOC model showed performance comparable to that of CA125 in serous carcinoma while demonstrating higher discrimination for clear cell, mucinous, and endometrioid carcinomas (Fig.  4 F). Fig. 4 Diagnostic performance of the two-step LightGBM-based model (CSGSA EOC model). (A) Risk score distributions (left) and ROC curves (right) for distinguishing EOC from all control groups (healthy individuals, uterine myoma, ovarian cysts, and endometrioma). (B) Score distributions and ROC curves by clinical stage (Stages I–IV) versus non-EOC controls. (C) Score distributions and ROC curves for each EOC histological subtype (clear cell carcinoma, mucinous carcinoma, endometrioid carcinoma, serous carcinoma) versus non-EOC controls. (D–F) ROC-AUC comparisons between the CSGSA model and conventional tumor markers (CA125, HE4, CA72-4) for (D) all EOC cases, (E) each stage, and (F) each histological subtype Diagnostic performance of the two-step LightGBM-based model (CSGSA EOC model). (A) Risk score distributions (left) and ROC curves (right) for distinguishing EOC from all control groups (healthy individuals, uterine myoma, ovarian cysts, and endometrioma). (B) Score distributions and ROC curves by clinical stage (Stages I–IV) versus non-EOC controls. (C) Score distributions and ROC curves for each EOC histological subtype (clear cell carcinoma, mucinous carcinoma, endometrioid carcinoma, serous carcinoma) versus non-EOC controls. (D–F) ROC-AUC comparisons between the CSGSA model and conventional tumor markers (CA125, HE4, CA72-4) for (D) all EOC cases, (E) each stage, and (F) each histological subtype Accurately distinguishing clear cell carcinoma from EM is particularly important, as a subset of EM cases may progress to clear cell carcinoma if left untreated [ 23 , 24 ]. We first evaluated the ability of three conventional tumor markers to distinguish between these two conditions. Both CA125 and HE4 showed limited classification performance, with ROC-AUC values below 0.6. CA72-4 achieved a ROC-AUC of 0.689, indicating weak discriminatory ability (Fig. 5 A). To improve discrimination, we developed a new machine learning model specifically designed to classify clear cell carcinoma versus EM, referred to as the “CSGSA CC model”. The CC model was developed using 120 clear cell carcinoma and 43 EM cases in the training cohort and evaluated in an independent hold-out test cohort consisting of 58 clear cell carcinoma and 23 EM cases. Given the limited sample size, the two-step modeling approach used in Model 3 was not applied. Instead, a LightGBM-based Model 2 was constructed using CA125, HE4, CA72-4, patient age, and 100 PCs derived from EGPs. Model development and prediction followed the same evaluation framework used for the CSGSA EOC model. The dataset was randomly divided into a training set (66%) and a hold-out test set (34%). Model optimization was performed exclusively using the training set, and performance was subsequently evaluated on the independent hold-out test set to prevent data leakage. The model achieved an ROC-AUC of 0.856, indicating improved discrimination between clear cell carcinoma and EM (Fig. 5 B). Feature importance analysis based on the Gain metric suggested that CA125 and HE4 contributed relatively little to the model, whereas PC2 and PC8 contributed substantially to classification (Fig. 5 C). Consistent with these findings, SHAP analysis identified patient age and PC2 as the major contributors to the model’s predictions (Fig. 5 D). The PCA score plot of PC2 and PC8 showed clear separation between clear cell carcinoma and EM along the PC2 axis (Fig. 5 E). Statistical analysis using Student’s t-test confirmed that differences in PC2 and PC8 scores between the two groups were highly significant ( p = 2.0 × 10 − 15 and p = 0.0039, respectively; Fig. 5 F and G). Fig. 5 Development of a machine learning model to distinguish clear cell carcinoma from endometrioma. (A) Risk score distributions (left) and ROC curves (right) for distinguishing clear cell carcinoma (CC) from endometrioma (EM) using conventional tumor markers (CA125, HE4, and CA72-4) and a newly developed subtype-specific model (CSGSA CC model). (B) Comparison of ROC-AUC values across the three markers and CSGSA CC model. (C) Feature importance (gain values) of the LightGBM-based CC model. (D) SHAP summary plot illustrating the feature contributions in the CSGSA CC model for discriminating between CC and EM. The plot displays the impact of each feature on the model output across all samples ranked by the mean absolute SHAP value. (E) PCA score plot showing sample distribution based on PC2 and PC8. Each dot represents a sample: red represents CC and yellow represents EM. (F , G) Box plots of the (F) PC2 and (G) PC8 scores by group. Statistically significant differences between CC and EM were assessed using the Student’s t-test Development of a machine learning model to distinguish clear cell carcinoma from endometrioma. (A) Risk score distributions (left) and ROC curves (right) for distinguishing clear cell carcinoma (CC) from endometrioma (EM) using conventional tumor markers (CA125, HE4, and CA72-4) and a newly developed subtype-specific model (CSGSA CC model). (B) Comparison of ROC-AUC values across the three markers and CSGSA CC model. (C) Feature importance (gain values) of the LightGBM-based CC model. (D) SHAP summary plot illustrating the feature contributions in the CSGSA CC model for discriminating between CC and EM. The plot displays the impact of each feature on the model output across all samples ranked by the mean absolute SHAP value. (E) PCA score plot showing sample distribution based on PC2 and PC8. Each dot represents a sample: red represents CC and yellow represents EM. (F , G) Box plots of the (F) PC2 and (G) PC8 scores by group. Statistically significant differences between CC and EM were assessed using the Student’s t-test

Epithelial ovarian cancer (EOC) is one of the leading causes of cancer-related mortality among women worldwide [ 1 , 2 ]. Early-stage EOC is frequently asymptomatic, leading to delayed diagnosis. Consequently, many patients are diagnosed only after the disease has progressed to advanced stages, when treatment options are more limited, and the prognosis is considerably poorer. In contrast, the 5-year survival rate for patients with early-stage EOC exceeds 90%, underscoring the importance of detecting the disease at a more treatable stage [ 3 ]. Current diagnostic approaches rely on imaging modalities such as transvaginal ultrasonography and computed tomography; however, their sensitivity for detecting early-stage disease in population-based screening settings remains limited [ 4 , 5 ]. Blood-based tumor markers, including cancer antigen 125 (CA125) [ 6 ] and human epididymis protein 4 (HE4) [ 7 ], have been widely investigated as potential screening tools. Nevertheless, their diagnostic performance, particularly in early-stage EOC, remains suboptimal [ 8 ]. In addition, the United Kingdom Collaborative Trial of Ovarian Cancer Screening (UKCTOCS) further demonstrated that screening strategies based on ultrasound or multimodal approaches did not result in a significant reduction in ovarian cancer-related mortality [ 9 , 10 ]. These findings highlight the limitations of existing imaging- and single-marker–based strategies and underscore the need for improved biomarker-based approaches for early detection [ 11 ]. Recent studies have further emphasized that cancer diagnosis and clinical decision-making increasingly benefit from the integration of complex molecular information, highlighting the growing importance of data-driven and biomarker-based approaches in oncology [ 12 ]. In this context, integrating multiple biomarkers with machine learning techniques offers a promising strategy for improving diagnostic performance [ 13 , 14 ]. In our previous work, we developed a technique termed comprehensive serum glycopeptide spectrum analysis (CSGSA), which analyzes enriched glycopeptides (EGPs) in serum using mass spectrometry to capture both glycan alterations and changes in protein expression [ 15 , 16 ]. This approach leverages the high specificity of existing tumor markers while incorporating glycan alterations associated with early-stage cancers, enabling evidence-based early detection. Mass spectrometry data were transformed into two-dimensional barcode-like images and analyzed using a convolutional neural network (CNN) to predict EOC risk [ 15 ]. This image-based framework allowed effective modeling with limited sample sizes through transfer learning. However, it requires labor-intensive preprocessing and may not fully capture key discriminative features of EOC. To address these limitations, we investigated alternative machine learning strategies better suited for structured tabular biomarker data. Gradient boosting algorithms, such as XGBoost [ 17 ] and LightGBM [ 18 ], have demonstrated high accuracy, efficiency, and robustness in biomedical applications, particularly with moderate sample sizes. These methods have been successfully applied to disease classification using metabolomic data, including lung cancer detection [ 17 ], and the diagnosis of type II diabetes [ 19 ] and diabetic retinopathy [ 20 ]. Together, these findings suggest that gradient boosting approaches may provide a practical and effective alternative to deep learning methods for biomarker-based classification. Beyond overall classification accuracy, improving the positive predictive value (PPV) is particularly important in ovarian cancer screening. Even with a specificity of 90%, the PPV remains below 1% when disease prevalence is less than 100 cases per 100,000 individuals, meaning that most individuals with positive results would not actually have cancer. This inherent limitation underscores the need for strategies specifically designed to improve the PPV in screening settings. To address these issues and enhance diagnostic performance, we adopted two complementary strategies. First, we developed a two-step diagnostic model based on gradient boosting that reflects clinical decision-making. In this framework, tumor marker levels and age are used in the initial step to stratify individuals into positive, negative, or indeterminate groups, followed by reclassification of the indeterminate group in a second step using additional features, including EGP-derived biomarkers. Second, we evaluated the addition of cancer antigen 72 − 4 (CA72-4), a tumor marker reported to be particularly useful for detecting mucinous carcinoma [ 21 ], a histological subtype of EOC that is often difficult to identify using conventional markers. Together, these strategies are designed to improve the PPV of the CSGSA framework and assess its potential as a blood-based approach for early detection of EOC.

In this study, we developed and evaluated a two-step LightGBM-based diagnostic model for early detection of EOC that integrates conventional serum tumor markers (CA125, HE4, and CA72-4), patient age, and EGPs. This approach achieved higher diagnostic performance than single-step or single-biomarker strategies, particularly for early-stage disease and for histological subtypes that are difficult to detect using conventional biomarkers. These findings are consistent with previous studies demonstrating that combining multiple biomarkers with machine learning methods can improve diagnostic accuracy in oncology. Notably, the model also improved PPV under low-prevalence screening conditions, addressing a key limitation of existing approaches. Several diagnostic algorithms have previously been proposed to improve ovarian cancer detection. For example, the Risk of Ovarian Malignancy Algorithm (ROMA) combines CA125 and HE4 to enhance detection beyond CA125 alone [ 25 ]. In addition, multi-marker panels and multi-analyte tests, such as those reported by Kang et al. [ 26 ], and CancerSEEK [ 27 ] have demonstrated promising diagnostic performance in multiple cancer types. Furthermore, liquid biopsy approaches, including circulating tumor cell detection, have emerged as valuable tools for early cancer detection and disease monitoring [ 28 ]. Our approach extends these strategies by incorporating glycopeptide-derived features and by adopting a two-step decision framework. Similar multi-marker machine learning approaches have also achieved high diagnostic accuracy in other cancers, including pancreatic and colorectal cancers (AUCs of 0.94–0.95) [ 29 – 32 ]. Together, these findings support the potential value of integrative biomarker approaches for improving cancer detection. An important feature of our approach is the hierarchical two-step diagnostic framework. This design is conceptually similar to established ovarian cancer screening approaches, such as the Risk of Ovarian Cancer Algorithm (ROCA), in which longitudinal CA125 measurements are used to estimate risk, followed by ultrasound and, if indicated, surgical intervention, achieving a PPV exceeding 10% while minimizing unnecessary surgical procedures to 2–3 per detected cancer [ 33 ]. Our model adopts a similar hierarchical framework; widely used serum tumor markers are applied in the first step to stratify risk, whereas enriched glycopeptide profiles are selectively incorporated in cases classified as indeterminate. A major objective of such hierarchical screening strategies is to improve PPV, which remains a major challenge in population-based ovarian cancer screening because of the low prevalence of the disease. Previous studies have suggested that a screening test would require a specificity of approximately 99.6% to achieve a PPV of approximately 10% in average-risk populations [ 34 ]. Even under such stringent conditions, many positive test results may still represent false positives, underscoring the fundamental challenge of ovarian cancer screening and the need for exceptionally high specificity. The two-step framework proposed in this study may help address this challenge by applying multiple decision stages, which can reduce the likelihood that non-cancer cases are classified as positive. Importantly, the rationale for the two-step framework extends beyond simply improving overall classification performance. Rather, it was designed to reflect a clinically relevant diagnostic workflow. In routine practice, patients with clearly benign findings or highly suspicious disease can often be classified using readily available clinical information, including age and serum tumor markers. Additional testing is most valuable for patients whose risk remains uncertain after this initial assessment. Consistent with this concept, narrowing the indeterminate range reduced the number of patients receiving second-stage glycopeptide evaluation; however, it also substantially reduced overall diagnostic performance. These findings suggest that applying glycopeptide analysis to diagnostically ambiguous cases may provide clinically meaningful incremental information beyond conventional tumor markers alone. This strategy resembles a stepwise diagnostic approach in which advanced testing is reserved for cases that remain difficult to classify using conventional clinical markers. Consistent with this diagnostic concept, the proposed two-step framework improved diagnostic efficiency and yielded a higher PPV (18.7%) than our previous single-step model [ 15 ]. Moreover, because CA125, HE4, and CA72-4 are widely used in clinical practice, this framework may facilitate integration into existing diagnostic workflows. Importantly, the proposed two-step framework also outperformed a conventional logistic regression model based solely on age and established ovarian cancer biomarkers (CA125, HE4, and CA72-4). Although both approaches achieved similarly high specificity, the two-step framework substantially improved sensitivity, AUC, and projected PPV. These findings suggest that the incorporation of glycopeptide information provides clinically meaningful diagnostic data beyond conventional tumor marker-based prediction alone. Nevertheless, interpretation of the projected PPV requires caution because it depends on assumed disease prevalence. Sensitivity analysis demonstrated that the projected PPV varied from 7.4% to 42.8% across the prevalence range evaluated (Additional File Table S7 ). As expected, the projected PPV was sensitive to disease prevalence, which is an inherent limitation of prevalence-adjusted estimates derived using Bayes’ theorem (Additional File Table S8 ). Therefore, the projected PPV of 18.7% should be interpreted as an estimate under predefined population assumptions rather than as a directly observed screening outcome. Prospective validation in real-world screening populations is required to establish the true PPV of the proposed framework. The performance improvement observed in the two-step framework was particularly pronounced when LightGBM was used as the underlying algorithm. Compared with XGBoost, LightGBM showed greater performance gains when additional glycopeptide-derived features were introduced in the second step, whereas only marginal gains were achieved with XGBoost. This difference may be related to the distinct tree-building strategies used by the two algorithms. LightGBM employs a leaf-wise tree growth strategy that prioritizes the most informative splits, enabling the efficient modeling of complex nonlinear patterns. This property is particularly advantageous in the second step of our framework, where enriched glycopeptide-derived features are introduced to resolve diagnostically ambiguous cases. In contrast, XGBoost adopts a level-wise tree growth strategy that favors balanced trees and robust generalization in relatively low-dimensional or homogeneous feature spaces but may be less effective in exploiting the high-dimensional and heterogeneous nature of omics-derived inputs. Consequently, XGBoost exhibits limited performance gains within a two-step framework. Together, these findings suggest that LightGBM may be well suited for modeling hierarchical, multilayered diagnostic frameworks that integrate conventional clinical markers with complex omics features. However, further evaluation across independent datasets is necessary to determine whether this advantage generalizes to other diagnostic applications. Despite the encouraging results of this study, several important limitations should be considered. First, although the hold-out test cohort was separated from the training data and was not used during model optimization, it was derived from the same overall multicenter dataset and therefore represents internal rather than true external validation. Consequently, the generalizability of the model to independent prospective populations, different healthcare settings, ethnic groups, and analytical environments remains uncertain. The overall dataset included both Asian and Caucasian participants, with both ethnic groups represented in the training and hold-out test sets after random splitting (Additional File Table S9 ). However, the study was not designed or powered to perform ethnicity-stratified performance analyses. Therefore, further validation in larger, ethnically diverse prospective cohorts is required to determine whether model performance is consistent across populations. Second, the performance metrics in real-world screening populations, including prevalence-adjusted PPV, may differ substantially from the estimates reported here. Even a projected PPV of 18.7% implies that a substantial proportion of positive screening results would remain false positives, which may result in additional clinical evaluations and psychological burden. Although the projected PPV remains imperfect, it compares favorably with those reported for many existing ovarian cancer screening approaches. Furthermore, as this study computationally builds upon a previously reported dataset collected between 2014 and 2022, it inevitably carries over any inherent temporal biases or pre-analytical limitations specific to that original collection period. This reliance on retrospective data further supports the interpretation of the present study as a proof-of-concept investigation. To further evaluate the robustness and generalizability of the proposed framework, a prospective blinded multicenter validation study involving more than 1,000 participants and independent sample collection is currently underway through the Japan Agency for Medical Research and Development project (AMED; JP24ama221416h0003) [ 35 ]. Third, although the inclusion of CA72-4 was associated with an increase in projected PPV (18.7% vs. 4.4%) despite similar ROC-AUC values (0.962 vs. 0.957), the statistical robustness of this observation remains to be established. Fixed-sensitivity analysis demonstrated only modest differences in specificity between models with and without CA72-4, and these differences did not reach statistical significance (Additional File Table S10 ). The apparent improvement in projected PPV may therefore reflect small shifts in specificity, which can disproportionately influence PPV under low-prevalence conditions. Accordingly, the contribution of CA72-4 should be interpreted cautiously and may represent a complementary component of the overall framework rather than a dominant independent driver of diagnostic performance. The clinical significance of this contribution should be confirmed in future studies. Fourth, the interpretability of the second-stage model may be limited. Although PCA effectively reduces the dimensionality of EGP data, it transforms them into abstract components that can obscure the contributions of individual glycopeptides and reduce biological transparency. While SHAP analysis was useful for identifying influential latent features within the model, the PCs themselves do not directly correspond to biologically interpretable entities. Therefore, biological interpretation relies primarily on the downstream analysis of glycopeptides majorly contributing to key PCs rather than on the PCs themselves. Several glycopeptides majorly contributing to the key PCs originated from transferrin, α1-acid glycoprotein, and haptoglobin. Reduced abundance of unsialylated transferrin glycoforms has been observed across multiple cancer types and may reflect cancer-associated hypersialylation resulting from systemic inflammatory responses and altered glycosyltransferase activity. In contrast, α1-acid glycoprotein and haptoglobin are acute-phase proteins known to exhibit increased branching, sialylation, and fucosylation during cancer progression. These glycosylation changes are thought to be driven by inflammatory cytokine signaling and altered hepatic glycan biosynthetic pathways. Therefore, although the PCs themselves are mathematical constructs, they may capture biologically meaningful alterations in cancer-associated glycosylation rather than merely representing abstract statistical features. The number of retained PCs was selected based on preliminary analyses performed within the training dataset. Although sensitivity analysis demonstrated that the model performance plateaued once approximately 100 PCs were included and was not highly sensitive to the exact number retained (Additional File Table S4 ), this approach may still introduce optimistic bias. A more rigorous strategy would incorporate the number of retained PCs as a hyperparameter within a nested cross-validation framework. Fifth, practical implementation considerations should also be acknowledged. In the current workflow, glycopeptide profiling requires approximately 30 min of LC–MS analysis per sample. A single LC–MS instrument can currently process approximately 50 samples per day, and throughput can be increased by deploying multiple instruments in parallel. Although these characteristics suggest that the proposed framework may be technically feasible for large-scale screening, the current implementation remains primarily a research-stage approach and requires specialized instrumentation and analytical expertise. Future efforts will focus on automating sample preparation and shortening LC–MS acquisition time from approximately 30 min to 10 min per sample, which could substantially improve throughput and reduce operational burden. Continued technological advances are also expected to improve analytical efficiency and cost-effectiveness. Further studies will be required to evaluate the real-world clinical utility, scalability, and economic feasibility of this approach in population-based screening settings. Sixth, the clear cell carcinoma sub-analysis was performed using a relatively limited number of EM cases. Although independent hold-out validation was performed, the possibility of overfitting cannot be completely excluded. The robustness of the CC model should therefore be interpreted cautiously until additional data become available. Seventh, samples were collected from 13 participating institutions. Although the same multicenter cohort was previously evaluated for inter-facility variability and no major institution-specific bias was observed, residual site-related effects cannot be completely excluded. Additional studies using independent cohorts will be important to further assess model robustness and generalizability. Eighth, to minimize potential confounding effects on serum glycopeptide profiles and establish an initial proof-of-concept model, patients with multiple cancers or ongoing hormonal therapy were excluded. Therefore, the study population may not fully reflect the complexity of real-world screening populations. These limitations highlight the need for further validation, methodological refinement, mechanistic investigations, and prospective clinical evaluation before broader implementation. Despite these limitations, several factors may facilitate future clinical implementation of the proposed framework. In Japan, transvaginal ultrasonography is widely available and routinely performed in gynecologic practice, allowing rapid exclusion of ovarian tumors and longitudinal follow-up of individuals with abnormal screening results. Such an infrastructure may facilitate future implementation of the proposed screening strategy and help mitigate the psychological impact associated with false-positive results. Most importantly, the ultimate goal of early detection strategies is to improve survival. Earlier detection of EOC can substantially increase 5-year survival rates, rising from less than 30% in advanced disease to 70–90% when diagnosed at early stages. Modeling studies further suggest that stage shifts in diagnosis could increase cure rates by 10–30% [ 36 ]. If successfully translated into prospective screening or surveillance settings, the proposed two-step model may contribute to achieving such a shift. In addition, several EGPs identified here may have biological significance beyond diagnostic value. Because glycopeptides reflect tumor-associated glycosylation changes, further investigation—especially in collaboration with proteomics and glycobiology experts—may clarify their biological roles in tumor progression or immune evasion, potentially opening new avenues for therapeutic or preventive strategies.

This study demonstrates that a two-step LightGBM-based model integrating serum tumor markers and enriched glycopeptide profiles achieved promising diagnostic performance for EOC detection in an internally validated cohort. Although the framework yielded an improved prevalence-adjusted PPV under the assumptions examined, prospective validation in independent external cohorts will be necessary to determine its real-world clinical utility and generalizability before clinical implementation can be considered.

Below is the link to the electronic supplementary material. Supplementary Material 1: Additional file 1: Supplementary Information. Table S1. Two-step LightGBM-based classification algorithm for training and prediction. Table S2. Area under the curve (AUC) and 95% confidence intervals of ROC analysis. (A) Comparison between epithelial ovarian cancer (EOC) and benign gynecological diseases. (B) Comparison between each stage of EOC and non-EOC. (C) Comparison between histological subtypes of EOC and non-EOC patients. (D) Comparison between clear cell carcinoma and endometriosis. Table S3. Effect of different indeterminate threshold ranges on the performance of the two-step LightGBM framework. Table S4. Effect of the number of retained principal components (PCs) on model performance. Table S5. Diagnostic performance of the proposed two-step LightGBM framework compared with a conventional biomarker-based logistic regression model. Table S6. Diagnostic performance of the two-step LightGBM model for epithelial ovarian cancer detection without CA72-4. (A) Confusion matrix. (B) Performance metrics. Table S7. Sensitivity analysis of the projected positive predictive value under different assumed prevalences of epithelial ovarian cancer. Table S8. Sensitivity of the projected positive predictive value to small changes in specificity. Table S9. Ethnic composition of the training and hold-out cohorts. Table S10. Comparison of specificity at a fixed sensitivity of 90% for models with and without CA72-4. Figure S1. SHAP waterfall plots for representative cases in the hold-out test cohort, including a false-positive case, true-negative case, true-positive case, and false-negative case Supplementary Material 1: Additional file 1: Supplementary Information. Table S1. Two-step LightGBM-based classification algorithm for training and prediction. Table S2. Area under the curve (AUC) and 95% confidence intervals of ROC analysis. (A) Comparison between epithelial ovarian cancer (EOC) and benign gynecological diseases. (B) Comparison between each stage of EOC and non-EOC. (C) Comparison between histological subtypes of EOC and non-EOC patients. (D) Comparison between clear cell carcinoma and endometriosis. Table S3. Effect of different indeterminate threshold ranges on the performance of the two-step LightGBM framework. Table S4. Effect of the number of retained principal components (PCs) on model performance. Table S5. Diagnostic performance of the proposed two-step LightGBM framework compared with a conventional biomarker-based logistic regression model. Table S6. Diagnostic performance of the two-step LightGBM model for epithelial ovarian cancer detection without CA72-4. (A) Confusion matrix. (B) Performance metrics. Table S7. Sensitivity analysis of the projected positive predictive value under different assumed prevalences of epithelial ovarian cancer. Table S8. Sensitivity of the projected positive predictive value to small changes in specificity. Table S9. Ethnic composition of the training and hold-out cohorts. Table S10. Comparison of specificity at a fixed sensitivity of 90% for models with and without CA72-4. Figure S1. SHAP waterfall plots for representative cases in the hold-out test cohort, including a false-positive case, true-negative case, true-positive case, and false-negative case Supplementary Material 2: Additional file 2: Source code for the two-step LightGBM framework (TwoSteps_LightGBM.py) Supplementary Material 2: Additional file 2: Source code for the two-step LightGBM framework (TwoSteps_LightGBM.py) Supplementary Material 3: Additional file 3: List of 1,712 enriched glycopeptide (EGP) features used for model development, including m/z and retention time information (EGP Peak List.xlsx) Supplementary Material 3: Additional file 3: List of 1,712 enriched glycopeptide (EGP) features used for model development, including m/z and retention time information (EGP Peak List.xlsx)