Nopour et al. (5) used a logistic regression (LR) model and an eXtreme Gradient Boosting (XGBoost) model for OC prediction and reported that XGBoost exhibited better accuracy than the regression model. Li et al. (6) adopted a logistic regression (LR)-based radiomics model out of other omics models (support vector machine (SVM), K-nearest neighbors (KNNs), Random Forest (RF), and XGBoost) and achieved an accuracy of 84.8%. Kori et al. (7) built a model using principal component analysis and ML algorithms that showed 96.7% sensitivity and 100% specificity. Klein et al. (8) built a model using MALDI imaging and ML algorithms such as linear discriminant analysis (LDA), SVM-lin, and SVM-rbf for epithelial ovarian cancer (EOC) histotype classification, yielding a mean accuracy of 80% for LDA, 80% for SVM-lin, and 74% for SVM-rbf. Chao et al. (9) designed an LR model for OC prediction and obtained the area under the curve (AUC), sensitivity, and specificity values of 0.903, 89.2, and 82.3% in the training dataset and 0.891, 88.9, and 76.75% in the test dataset, respectively. Schilling et al. (10) used ML models using K-means clustering, Naive Bayes, logistic regression, and SVM for disease prediction and revealed that the LR model performed the best with a 0.89 f1 score. Hong et al. (11) built a radiomics model based on LASSO Cox regression and reported good discrimination in both training and validation sets with C-indexes of 0.694 (95% confidence interval [CI]: 0.613–0.775) and 0.709 (95% CI: 0.517–0.901), respectively. Bahado-Singh et al. (12) worked with ML algorithms (SVM, LDA, prediction analysis for microarrays (PAMs), and generalized linear model (GLM)) and obtained optimal results of AUC value (95% CI) 1.00 (0.9000–1.0) with SVM and AUC (95% CI) 0.99 (0.9000–1.0) with GLM, PAMs, and LDA, along with achieving an AUC value (95% CI) 1.00 (0.9–1.0) with SVM models for CpG analysis. Sheela Lavanya et al. (13) applied ML techniques (KNNs, SVM, DTs) alongside explainable AI to aid the early detection of OC and demonstrated that SVM performed with 85% base model accuracy. Feng Zhan et al. (14) designed the LR radiomic model to predict lymphocyte-specific protein tyrosine kinase (LCK) expression and overall survival in high-grade serous ovarian cancer (HGSOC) patients and achieved AUC values of 0.879 and 0.834, respectively. A.D Coles et al. (15) performed a comparative analysis of the RF and LR models for retrospective detection of OC recurrences from chemotherapy data and revealed that RF achieved the highest F1 score among the two. Alexander Laios et al. (16) performed ML-based risk prediction using the KNN and LDA techniques and QDA algorithms that achieved predictive accuracies of 0.80, 0.90, and 0.92, respectively.

Subsequent studies used the following ensemble models for the prediction of OC intrinsically: Random Forest (RF), XGBoost, gradient-boosted trees (GBT), and Bagging. Gong et al. (17) used nine supervised ML classifiers including RF and GBT, out of which RF performed with high stability and an accuracy of 0.60 including gut microbiota for chemoresistance to OC prediction using the RF model with an AUC value of 0.909. Ahamad et al. (18) utilized a combination of XGBoost, GBT, and light gradient boosting machine (LGBM) to build classifier models, and the predictive analysis results revealed an accuracy of 91%. Hamidi et al. (19) developed Boruta, a novel RF feature selection-based ML model, for identifying important biomarkers. Zhao et al. (20) used LASSO regression, ridge regression, XGBoost, RF, and AdaBoost to build a LASSO Cox regression model. By performing decision curve analysis, the resulting nomogram indicated that the combined model for predicting 1-year and 3-year survival probabilities provided an optimal net benefit compared to using a single indicator. Wadapurkar et al. (21) applied next-generation sequencing analysis to cancer driver genes and OC prediction using the XGBoost classification model and yielded an accuracy of 0.946. Abuzinadah et al. (22) used the RF model, GBM, and also ensemble GBM + XGB models. GBM was reported to be the best with an accuracy of 87.14%, a recall of 87.53%, and a precision of 87.58%. Zeng et al. (23) used a new multi-omics model including all the features and showed the best prediction performance with an AUC value of 0.911. Piedimonte et al. (24) used a Random Forest model that was trained to predict the dichotomous outcome of optimal cytoreduction and obtained an AUC of 99.8%. Cheng et al. (25) used a gradient-boosting decision tree algorithm in which the AUC value of the signature gene pair was 0.9658, whereas the AUC value of the individual signature gene-based prediction was only 0.6823. Seri Jeong et al. (26) utilized the RF model for differentiating benign and malignant OC using combined cancer markers and achieved receiver operating characteristic (ROC) AUC values of 0.707, 0.680, 0.643, 0.657, and 0.624 for the biomarkers ROMA, HE4, CA125, LD, and NLR, respectively. Maria et al. (27) used an XAI-based ensemble model for the accurate classification of ovarian cancer and achieved an accuracy of 83.2%. Additionally, the model proved to be highly effective in handling imbalanced datasets. Annarita et al. (28) proposed a waterfall-based classification model for the classification of clinical data and ultrasound indicators and achieved an accuracy of 86.36%.

Wu et al. (29) used an attention-based network by using a pre-trained model ResNet50 on ImageNet to obtain a C-index value of 0.5789 (0.5096–0.6053) and a p-value of 0.00845, while the risk score indicated good prediction ability in the homologous recombination deficiency (HRD+) subgroup. Feng et al. (30) used a backpropagation neural network model by using a pre-trained ResNet50 which outperformed the U-net-based network architecture by obtaining 89.11% average sensitivity and 96.37% specificity. Suganya et al. (31) used MALDI imaging for the EOC classification. Out of several ML and DL models used in the research, neural network and convolutional neural network (CNN) were most suitable for EOC classification with an accuracy of 83% for NN, 85% for CNN, and a sensitivity of 69–100% as well as specificity of 90–99% for both the models, respectively. Rahul Mishra et al. (32) used the Convolutional Neural Networks with Gray Wolf Optimization (CNNGWO) model and reported an improved diagnostic accuracy of OC to 98%, while Talib et al. (33) trained a simple CNN-based DL model on 1798 images and achieved an accuracy of 81% and an AUC value of 0.89. Kasture et al. (34) used outclassed traditional CNN algorithms, deep CNN, and pre-trained AlexNet models, improving the accuracy from 70 to 89.93%. Reilly et al. (35) proposed an MIA3G model (deep feed-forward neural network) for OC risk assessment and achieved a sensitivity of 89.8%, specificity of 84.02%, positive predictive value (PPV) of 22.455, and negative predictive value (NPV) of 99.38%. David Joon Ho et al. (36) worked with a deep multi-magnification network model (DMMN) for OC segmentation and obtained 86% recall and 84% precision. Ziyambe et al. (37) built a CNN architecture model using the Xception network and obtained 94.43% accuracy, 95.02% sensitivity, and 93.16% specificity, making it a potential tool in assisting physicians for predicting OC. Dingdu Hu et al. (38) conducted automatic segmentations of T2-weighted MRI using DL algorithms (CNN, U-Net, DeepLabv3, U-Net++, and PSPNet) along with transformers (TransUnet and Swin-Unet), and reported that U-Net++ performed the best with precision and recall rates of 83.8 and 88.2% in the internal test set and 82.5 and 72.5% in the external test set, respectively. Ravishankar et al. (39) proposed a fuzzy CNN-based classifier with an initial accuracy of 98.37%, which later achieved 100% accuracy occasionally with an increase in several epochs. Wang et al. (40) used two improved DL-based methods for predicting bevacizumab efficacy in EOC patients based on histopathological images and revealed that the proposed model using the MSH2 protein expression performed excellently with 100% accuracy, sensitivity, and specificity, respectively. Ju Young et al. (41) developed a U-Net-based DL algorithm for the characterization of cancer-associated thrombosis of OC and achieved outstanding results with an AUC value of 0.99 along with high sensitivity and specificity. Jeya Sundari et al. (42) developed an intelligent black widow optimization on image enhancement with a DL-based ovarian tumor diagnosis model, which achieved a maximum contrast of 0.97, a contrast-to-noise ratio (CNR) of 92.74%, a weighted peak signal-to-noise ratio (WPSNR) of 20.43, and a homogeneity of 0.94. Rui Yin et al. (43) worked with a multitask deep learning model for high-grade serous ovarian carcinoma (HGSOC) prediction using CT images of OC, which resulted in AUC values of 0.87 (0.83–0.89), 0.88 (0.86–0.91), 0.86 (0.82–0.89), and 0.79 (0.75–0.82) in the training sets, validation sets, prospective sets, and external sets, respectively. Yang et al. (44) developed the Ovarian Cancer Digital Pathology Index (OCDPI), a graph-based DL model, for prediction and treatment response in OC patients using histopathological images and demonstrated prognostic ability for overall survival prediction in the prostate, lung, colorectal, and ovarian cancer (PLCO; HR, 1.916; 95% CI, 1.380–2.660; log-rank test, p < 0.001) and Harbin Medical University Cancer hospital (HMUCH) (HR, 2.796; 95% CI, 1.404–5.568; log-rank test, p = 0.0022) cohorts. It was observed that the performance of OCDPI was good in patients with low-grade tumors or fresh frozen slides. Laios et al. (45) developed a predictive model using XGBoost and the deep neural network for the prediction of epithelial ovarian cancer and used a SHAP technique to discuss how intraoperative decisions during EOC cytoreduction can integrate human factors along with factual knowledge to optimize the selected trade-off in surgical effort. Huang et al. (4) emphasized classification tasks and used a game-theoretic framework using Shap values for model development, evaluation, and result visualization. The case study findings highlighted the effectiveness, reliability, and benefits of the ML pipeline in comparison to traditional statistical methods.

The dataset comprises 1,500 cases of ovarian tumors, classified into borderline, malignant, and benign types. The authors meticulously transferred the data manually from medical records to an Excel file, following ethical medical practices under the vigilance of medical practitioners at the Kasturba Medical College, Karnataka, India. Each row represents a unique patient and the columns capture various clinical features pertinent to the presence or absence of ovarian cancer. The raw data from the records were structured into the following 28 columns in the Excel file: clinical parameters such as age, parity, family history, oral contraception, age at menarche, menopause, abdominal pain, menstrual abnormalities, dysmenorrhea, abdominal distention, loss of appetite, edema, pallor, P/V tenderness, and abdominal mass; ultrasound or CT parameters: bilateral, ascites, solid area, papillary projection, thick septa, vascularity, contrast enhancement, lymph nodes, and deposits; and tumor markers such as CA125, CA19-9, and CEA and the corresponding category or class label. There are three classes: 0 – benign, 1 – border, and 2 – malignant.

Some of the feature descriptions are as follows:

This dataset served as a vital tool for studying the clinical features and diagnostic markers of ovarian tumors, aiding in advancing medical research and treatment strategies.

Figure 1 depicts the process flow followed in the current study. The dataset consists of 1,500 samples, with benign, malignant, and border cases comprising 500 samples each. The data are stored in a CSV file. This is considered the Input Data file that holds the raw data. Further steps were performed on this data. Preprocessing steps such as feature scaling were performed using a standard min–max scaler. A standard min–max scaler was applied to rescale all features to a uniform range, typically [0, 1]. This ensured that features with different scales did not disproportionately influence the model. Feature Transformation using exponential transformation was used to address skewed data distributions; the model was enhanced to capture patterns in the data along with labeling instances into correct categories, and missing values were imputed using mean imputation, replacing missing entries with the mean of the respective features. This approach preserved the overall distribution of the data and maintained dataset integrity without introducing significant bias. The processed data were then provided as input for several machine learning classifier models. The ML classifiers K-nearest neighbors (KNNs), logistic regression (LR), support vector machine (SVM), decision trees (DTs), and Naive Bayes (NB) were used in the current study. For the ensemble classifiers, models such as Random Forest (RF), XGBoost, CatBoost, and GradientBoost were applied. All classifiers were evaluated based on their performance using evaluation metrics such as accuracy, AUC, and ROC. The best-performing models among all the models were chosen. Random Forest, XGBoost, and SVM with accuracies of 98.13, 97, and 97.06%, respectively, were the top three best-performing models and were used to design a three-stage ensemble architecture. These scores represent the performance of the models on the test data. Explainable AI methods were applied to the test data results and passed onto the final ensemble model. LIME and SHAP were used to visualize and explain the predictions made by the model, as they are widely used for explanation across domains.

The results have been computed by considering the complete dataset of clinical parameters, tumor markers, and ultrasound parameters. The accuracy rates of single classifiers and ensemble classifiers are depicted as bar charts in Figure 2, respectively. It can be observed that SVM and Random Forest exhibit the highest accuracy among the single classifiers and the ensemble models, respectively. Only SVM, Random Forest (RF), and XGBoost have an accuracy of 97% or above, with 97.06, 98.13, and 97%, respectively. Therefore, these models were used to build the three-stage ensemble classifier. In Figure 3, the accuracy rates of all the models used in the current study have been plotted along with the proposed three-stage ensemble classifier.

Figure 4 shows the ROC curves for SVM, Random Forest, and the proposed three-stage ensemble models. Their AUC–ROC values are 0.92, 0.95, and 0.96, respectively. It is evident from the figure that the proposed model exhibits the best predictive behavior. In machine learning, feature importance is a technique that assigns a score to input features based on their importance for predicting the target variables. Since the proposed model is a combination of Random Forest, SVM, and XGBoost, it naturally provides the feature importance scores as they are the tree-based model. The RF model gives the feature importance based on the entropy score for the classification tree and aggregates the scores of each feature across all the trees in the forest. Each tree’s feature importance is calculated based on the reduction in impurity (entropy value). Similar to RF, GradientBoost calculates the feature importance score using the total reduction in the log loss. The feature importance predicted on the overall dataset and only on the clinical parameters is shown in Figures 5A,B, respectively.

Figure 5A shows the feature importance of each feature contributing to ovarian malignancy. It is evident that CEA, loss of appetite, CA19-9, CA125, abdominal distention, family history, age of menarche, and age are the highest contributors and more prominently influence the decision of which category the particular case will fall into. It can be seen from Figure 5B that the top features family history, loss of appetite, age, abdominal distention, oral contraception, and age of menarche are the highest contributors among clinical parameters and more prominently influence the decision of which category the particular case will fall into. In contrast, the features such as dysmenorrhea, menopause, menstrual abnormalities, and abdominal pain are the lowest contributors clinically, however, but are of significance when accounted for as an isolated event.