Machine-Learning-Assisted Analysis of Patient Clinical Biomarkers to Improve Ovarian Cancer Diagnosis.

All supervised machine learning analyses were performed on a standard laptop equipped with an Intel i7–7500U processor and 8 GB RAM and installed with Orange Data Mining software (ver. 3.36.2, University of Ljubljana, Slovenia). The computational cost for training all classifiers was minimal, generally less than 10 s per classifier, although gradient boosting and neural network required slightly longer than that due to their iterative nature of training. The data set used in this work was acquired from a previously published study, and it consists of data from 178 patients with cancerous ovarian tumors and 171 patients with benign ovarian tumors. The data set consists of 47 features, which are broadly categorized into blood cells, blood chemistry, and tumor biomarkers. The data set was subjected to preprocessing to remove entries with incomplete features prior to subsequent analysis, which resulted in an eventual data set consisting of 140 patients with cancerous tumors and 169 patients with benign tumors ( Supporting Information Excel File 1 ). Feature scoring and ranking, principal component analysis (PCA), and t -distributed stochastic neighbor embedding ( t -SNE) plotting were performed on Orange Data Mining software (ver. 3.36.2, University of Ljubljana, Slovenia). Features were scored based on information gain ratio and Gini index. A total of 8–10 best-ranked features were selected for further analysis. For PCA, the number of components was selected such that at least 80% of the data variance was explained. For visualization of data distribution through t -SNE, the perplexity was set at 40. All statistical analyses were performed using GraphPad Prism software (ver. 10.2, GraphPad Software, Inc.). The distribution of the data was first evaluated. Next, t test with Welch’s correction was employed to evaluate the parametric data, while the nonparametric data were analyzed using Mann–Whitney test. * p < 0.05, *** p < 0.001, and **** p < 0.0001 indicate statistically significant difference. ns indicates not statistically significant difference. The preprocessed data set was randomly split into 80% training and 20% testing sets, which were used to train and test all supervised machine learning algorithms, respectively. Seven supervised machine learning algorithms, i.e., logistic regression, random forest, gradient boosting, adaptive boosting, k -nearest neighbors, neural network, and support vector machine, were trained and tested on Orange Data Mining software (ver. 3.36.2, University of Ljubljana, Slovenia). Each algorithm was trained and tested against six distinct data sets, which encompass (i) all patients with all 47 features, (ii) all patients with 10 selected features, (iii) premenopausal patients with all 47 features, (iv) premenopausal patients with 10 selected features, (v) postmenopausal patients with all 47 features, and (vi) postmenopausal patients with 10 selected features. During the training phase, the hyperparameters of each algorithm were fine-tuned to optimize its predictive performance. A stratified 10-fold cross-validation was used to minimize overfitting and bias. The hyperparameters that generated the best algorithm performance for each data set were selected for testing. For logistic regression, (i, ii, iii, and vi) lasso regularization with a strength of (i and vi) 200, (ii) 1, and (iii) 1000, as well as (iv) and (v) ridge regularization with a strength of 1000 were employed. For random forest, the number of trees used was (i, ii, v, and vi) 50, (iii) 200, and (iv) 200. For gradient boosting, CatBoost method, 200 trees, a learning rate of 1, a regularization value of (i, ii, iii, v, and vi) 0.0001 and (iv) 0.01, and a limit 12 individual trees were selected. For adaptive boosting, 10 estimators, a learning rate of 0.01, the SAMME classification algorithm, and a linear regression loss function were utilized for all data sets. For k -nearest neighbors, 1 neighbor, the Euclidean metric, and distance-based weighting were used across all data sets. For the neural network, (i, iii, iv, v, and vi) 500 and (ii) 100 neurons in hidden layers, rectified linear unit (ReLu) activation function, the limited-memory quasi-Newton method-based optimizer (L-BFGS-B) solver, a regularization of (i, ii, and iv) 0.01 and (iii, v, and vi) 0.0001, and a maximum iteration limit of (i, iii, iv, v, and vi) 500 and (ii) 200. For support vector machine, a cost of (i and v) 10 and (ii, iii, iv, and vi) 100, a regression loss epsilon of (i, ii, iii, v, and vi) 0.1 and (iv) 1, a kernel function of (i, iii, and vi) radial basis function (RBF), (ii and iv) polynomial, and (v) linear, and an iteration limit of (i, ii, iii, iv, and vi) 10,000 and (v) 100 were chosen. The performance of all algorithms was evaluated based on quantitative metrics, notably, the accuracy, precision, recall, and F 1 score, which were computed based on eqs – : 1 accuracy = TP + TN TP + TN + FP + FN 2 precision = TP TP + FP 3 recall = TP TP + FN 4 F 1 = 2 × p recision × r ecall p recision + r ecall where TP is the number of true positives, that is, the number of correctly predicted positive instances (i.e., cancerous tumors); TN is the number of true negatives, that is, the number of correctly predicted negative instances (i.e., benign tumors); FP is the number of false positives, that is, the number of incorrectly predicted positive instances (i.e., benign tumors predicted as cancerous tumors); and FN is the number of false negatives, that is, the number of incorrectly predicted negative instances (i.e., cancerous tumors predicted as benign tumors).

The workflow of our study is summarized in Figure . Briefly, a publicly available data set comprising data on patients with cancerous ovarian tumors ( n = 178) and benign ovarian tumors ( n = 171) and with 47 features, which were broadly classified into blood cells, blood chemistry, and tumor biomarkers, was first obtained from a previously published study. This raw data set was then preprocessed to remove entries with incomplete features, resulting in a data set comprising 140 patients with cancerous tumors and 169 patients with benign tumors ( Supporting Information Excel File 1 ). The processed data was next subjected to feature scoring and ranking, based on information gain ratio and Gini index, to identify important features that could be used to differentiate between cancerous and benign tumors. Principal component analysis (PCA) was conducted to transform the high-dimensional patient data with 47 features into that with 16 to 18 principal components accounting for at least 80% of the data variance. This was followed by verification of data distribution in a two-dimensional space using t -distributed stochastic neighbor embedding ( t -SNE) plots. In parallel, the data set was randomly split into 80% training and 20% testing sets, which were used to train and test seven supervised machine learning algorithms (i.e., logistic regression, random forest, gradient boosting, adaptive boosting, k -nearest neighbors, neural network, and support vector machine), respectively. A stratified 10-fold cross-validation was implemented to tune the hyperparameters of the classifiers to yield the highest performance metrics. The classification performance of these algorithms in terms of accuracy, precision, recall, and F 1 value was evaluated using confusion matrices and receiver operating characteristic (ROC) curves. Overview of the workflow in analyzing patient ovarian tumors. A raw data set comprising data on patients with cancerous and benign ovarian tumors was first preprocessed. Feature scoring and ranking were then performed on the preprocessed data set to identify several important features that could be capitalized on to distinguish between cancerous and benign ovarian tumors. Principal component analysis was conducted to transform the high-dimensional data set into its low-dimensional counterpart, followed by visualization of data distribution in a two-dimensional space. Separately, the preprocessed data set was split into training and testing data sets to train and test numerous supervised machine learning algorithms, respectively, whose classification performance was evaluated through confusion matrices and receiver operating characteristic curves. While it is useful to analyze large data sets with many features, it is noteworthy that not all features hold equal importance in determining outcomes. This is especially true in the identification and use of predictive markers for cancer diagnosis. Motivated by this, the preprocessed patient data set was first subjected to feature scoring and ranking to determine essential factors which could be potentially used to differentiate between cancerous and benign ovarian tumors ( Figure ). Specifically, the information gain ratio and Gini index were selected as the evaluation metrics ( Figures a and S1 ). It is interesting to note that ovarian cancer biomarkers, i.e., HE4 and CA125, along with menopausal status and age emerged as the most important features in distinguishing cancerous and benign ovarian tumors. Alkaline phosphatase, albumin, lymphocyte ratio, lymphocyte count, and platelet count ranked next in terms of their importance. Analysis based on the Gini index showed that the same features identified through information gain ratio ranked as the most important features too ( Figure S1 ). PCA was subsequently performed to reduce the dimensionality of the data, followed by their visualization in a two-dimensional space using a t -distributed stochastic neighbor embedding ( t -SNE) plot ( Figure b). Statistical analysis of the eight most important features revealed that there were significant differences between these characteristic features of the cancerous and benign tumors ( p < 0.001) ( Figure c–j). Scoring and comparative analysis of various features of all patients with ovarian tumors. (a) Feature scoring and ranking based on information gain ratio. (b) t -SNE map illustrating the distribution of benign (BT) and cancerous (CT) ovarian tumors in two-dimensional space. (c, j) Comparative analysis of important differentiating features as identified through feature scoring: (c) HE4, (d) menopausal status (i.e., 0 indicates premenopause and 1 indicates postmenopause), (e) CA125, (f) age, (g) alkaline phosphatase, (h) albumin, (i) lymphocyte ratio, and (j) platelet count. n = 169 for benign tumors and 140 for cancerous tumors. *** p < 0.001 and **** p < 0.0001 based on Mann–Whitney test. Based on feature scoring and ranking, it appeared that menopausal status and age were crucial in the pathological development of ovarian tumors in all patients. To further evaluate the contribution of various features, the patients were split into two groups, i.e., premenopausal ( Figures and S2 ) and postmenopausal ( Figures and S3 ) groups, and feature scoring and ranking were repeated. Scoring and comparative analysis of various features of premenopausal patients with ovarian tumors. (a) Feature scoring and ranking based on information gain ratio. (b) t -SNE map illustrating the distribution of benign (BT) and cancerous (CT) ovarian tumors in two-dimensional space. (c–j) Comparative analysis of important differentiating features as identified through feature scoring: (c) HE4, (d) albumin, (e) CA125, (f) age, (g) red blood cell distribution width, (h) lymphocyte ratio, (i) mean corpuscular hemoglobin, and (j) carcinoembryonic antigen. n = 146 for benign tumors and 58 for cancerous tumors. * p < 0.05, *** p < 0.001, and **** p < 0.0001 based on either Mann–Whitney test or t test with Welch’s correction. Scoring and comparative analysis of various features of postmenopausal patients with ovarian tumors. (a) Feature scoring and ranking based on information gain ratio. (b) t -SNE map illustrating the distribution of benign (BT) and cancerous (CT) ovarian tumors in two-dimensional space. (c, j) Comparative analysis of important differentiating features as identified through feature scoring: (c) CA125, (d) HE4, (e) thrombocytocrit, (f) indirect bilirubin, (g) platelet count, (h) basophil ratio, (i) total bilirubin, and (j) albumin. n = 23 for benign tumors and 82 for cancerous tumors. * p < 0.05, *** p < 0.001, and **** p < 0.0001 based on Mann–Whitney test. Evaluation of feature importance in the premenopausal group uncovered that both HE4 and CA125 ranked as the most important features ( Figures a and S2 ). Along with these ovarian cancer biomarkers, albumin, age, and red blood cell distribution emerged as the top five most important features. Attributes like the lymphocyte ratio, mean corpuscular hemoglobin, and carcinoembryonic antigen also ranked highly in their importance in differentiating cancerous and benign ovarian tumors. Next, PCA was used to reduce the dimensionality of the data, and their distribution was visualized in a two-dimensional space through t -SNE, which showed a higher number of patients with benign tumors than cancerous tumors in the premenopausal group ( Figure b). Statistical analysis of the eight highly ranked features elucidated that benign and cancerous ovarian tumor patients exhibited statistically significantly different HE4, albumin, CA125, age, red blood cell distribution width, lymphocyte ratio, and mean corpuscular hemoglobin ( p < 0.05) ( Figure c–i). Intriguingly, the difference in the carcinoembryonic antigen levels between the benign and cancerous tumor groups was statistically negligible ( Figure j). Like the premenopausal group, for the postmenopausal group, both CA125 and HE4 still ranked as the most essential features distinguishing benign and cancerous ovarian tumor groups ( Figures a and S3 ). Thrombocytocrit, indirect bilirubin, platelet count, basophil ratio, total bilirubin, and albumin followed closely as the next important features. The dimensionality of the data was reduced through PCA, and t -SNE was employed to visualize data distribution, revealing a higher proportion of patients with cancerous tumors than benign tumors in the postmenopausal group ( Figure b). Quantitative analysis of several highly ranked features showed statistically significant differences between the attributes of the two tumor groups ( p < 0.05) ( Figure c–j). Nonetheless, although basophil ratio placed highly in feature scoring and ranking, the difference was statistically insignificant ( Figure h). One of the most important advantages of machine learning for cancer diagnosis is its ability to classify cancerous tumors from their benign counterparts, which can be potentially used to predict the early occurrence of cancer. As such, in this part of the study, we were motivated to evaluate whether supervised machine learning could be capitalized on to classify cancerous and benign ovarian tumors in different female patient populations. Specifically, we aimed to assess the predictive performance of different types of supervised machine learning algorithms, i.e., regression, tree-based, instance-based, and deep learning algorithms. To begin with, the data sets comprising all patient populations with all and selected features were first split into 80% training data sets and 20% testing data sets ( Figures , S4, and S5 ). Seven supervised machine learning algorithms, i.e., logistic regression, random forest, gradient boosting, adaptive boosting, k -nearest neighbors, neural network, and support vector machine, were then trained using the training data set. Throughout the training phase, all algorithm hyperparameters were fine-tuned to optimize the predictive performance of algorithms, and the hyperparameters yielding the best performance were selected for the testing phase. The performance of all algorithms was evaluated based on four quantitative metrics, i.e., accuracy, precision, recall, and F 1 score. Accuracy measures the frequency of correct outcomes as identified by the algorithm, while precision gauges the accuracy of the predicted positive data. Recall, also known as sensitivity, quantifies how often positive data are correctly identified among all positive instances, thereby assessing the completeness of positive predictions. The F 1 score, which is a harmonic mean of precision and recall, provides a comprehensive measure of the algorithm performance. Supervised classification of ovarian tumors in all patients. (a) Table summarizing the performance of supervised machine learning algorithms on testing data set taking all features into account. (b) Confusion matrices and (c) ROC curves of the algorithms with better performance (i.e., logistic regression and random forest). (d) Table summarizing the performance of supervised machine learning algorithms on testing data set taking only the most essential features into account. (e) Confusion matrices and (f) ROC curves of the algorithms with better performance (i.e., gradient boosting and random forest). We noted that all algorithms exhibited strong predictive performance during the training phase, with more than 80% classification accuracy, precision, recall, and the F 1 score. Among the algorithms, random forest and logistic regression were the top performers ( Figure S4 ). When evaluated against the testing data set, the predictive performance of all hyperparameter-tuned algorithms ranged from 76.7 to 90.3% across the assessed metrics ( Figure a–c). Particularly noteworthy are the logistic regression and random forest algorithms, which boasted an accuracy of 90.3% and precision of 90.9 and 90.4%, respectively ( Figure a). Of the 62 ovarian tumor patients in the testing data set, logistic regression correctly classified 33 patients as having benign tumors and 23 patients as having cancerous tumors, which translates to one false-positive and five false-negative values. In contrast, the random forest had two false positives and four false negatives in its predictions ( Figure b). Assessing the ROC curves, especially the area under the curve (AUC) values, which serve as a metric for assessing the effectiveness of each algorithm in discriminating between benign and cancerous tumors, we observed that random forest and logistic regression achieved considerably high AUC values of 91.4 and 91.1%, respectively. To examine if the number of features would influence the predictive performance of the supervised machine learning algorithms, we next assessed their performance metrics using only 10 most essential features as identified in the previous part of the study through feature scoring and ranking (i.e., information gain ratio) ( Figures d–f and S5 ). We noted that during the training phase, apart from logistic regression, k -nearest neighbors, and adaptive boosting, the rest of the algorithms showed an overall increase in their classification performance. Notably, the performance metrics for random forest and gradient boosting improved, with classification accuracy rising from 87.0 to 87.9% and from 85.8 to 88.3%, respectively ( Figures S4 and S5 ). Interestingly, except for logistic regression and k -nearest neighbors, most algorithms demonstrated a considerable improvement in performance metrics during the testing phase ( Figure d). Particularly noteworthy is the significant improvement shown by the gradient boosting algorithm, in which its accuracy, precision, recall, and F 1 score increased from about 80 to 90%. The improved classification performance of the different algorithms, as the number of features considered was reduced from 47 to 10, was also reflected in their confusion matrices and AUC values ( Figure e,f). Next, we were motivated to assess whether the performance of the supervised machine learning algorithms in classifying ovarian tumors would be influenced by the menopausal status of the patients. Consequently, we grouped all patients into premenopausal and postmenopausal patients and evaluated the predictive performance of the machine learning algorithms on each population. We first started with the data sets on premenopausal patients with all and selected features and split the data sets into 80% for training and 20% for testing ( Figures , S6, and S7 ). Evaluating the algorithm performance on the training data set containing all features, we noted that the four metrics of all algorithms were between 79 and 86.8% ( Figure S6 ). These metrics decreased substantially when the algorithms were assessed against the testing data set ( Figure a). Particularly, gradient boosting and logistic regression showed the best classification performance with 82.9% accuracy and 86.3% precision. It is worth mentioning that, as compared to the classification performance on the testing data set comprising all patients, the performance of most of the algorithms on the premenopausal patient testing data set was weaker. Examining the confusion matrices of the best-performing algorithms, we noted that those of gradient boosting and logistic regression contained seven false negatives, which represented more than half of the total number of cases of cancerous ovarian tumors in the premenopausal patient data set ( Figure b). Further assessment of the algorithm performance metrics unveiled that gradient boosting and logistic regression had AUC values of 80.5 and 76.9%, respectively ( Figure c). Supervised classification of ovarian tumors in premenopausal patients. (a) Table summarizing the performance of supervised machine learning algorithms on testing data set taking all features into account. (b) Confusion matrices and (c) ROC curves of the algorithms with better performance (i.e., gradient boosting and logistic regression). (d) Table summarizing the performance of supervised machine learning algorithms on testing data set taking only the most essential features into account. (e) Confusion matrices and (f) ROC curves of the algorithms with better performance (i.e., gradient boosting and neural network). The predictive performance of all algorithms on the premenopausal patient data set containing only the 10 most important features was subsequently evaluated ( Figures d–f and S7 ). Apart from logistic regression and k -nearest neighbors, all algorithms showed higher classification accuracy, precision, recall, and F 1 score during the training phase ( Figures S6 and S7 ). Evaluation of the algorithm classification performance against the testing data set revealed that gradient boosting, neural network, and adaptive boosting were the better-performing algorithms ( Figure d). Specifically, these three algorithms achieved classification accuracy, precision, recall, and F 1 score of 82.9, 83.6, 82.9, and 81.6%, respectively. The confusion matrices of the two best-performing algorithms showed that both gradient boosting and neural network classified 27 out of 28 benign tumors correctly, although these algorithms missed less than half of the total number of cases of cancerous tumors ( Figure e). Also, gradient boosting had an AUC value of 82.4%, while that of the neural network was 80.5% ( Figure f). Here, it is crucial to recognize that, for both premenopausal patient data sets, the false-negative rates of some of the best-performing algorithms were especially high. We noted that classifiers like gradient boosting, logistic regression, and neural network misclassified close to or more than half of the cancerous tumors as benign ( Figure b,e). This observation raises nontrivial concerns about the clinical applicability of these supervised classifiers, particularly if they were to be utilized to assess high-risk populations, where inaccurate diagnoses may delay timely intervention and negatively impact patient outcomes. Although the overall performance metrics of all classifiers were relatively promising, such subgroup-specific vulnerabilities underscore the need for further optimization. Consequently, more rigorous validations and targeted improvements to reduce the risk of false negatives must be made prior to the implementation of this supervised machine learning approach for clinical decision-making. After assessing the performance of all supervised machine learning algorithms on all and premenopausal patient data sets, we sought to evaluate the algorithm performance on the data sets consisting of postmenopausal patients with all and selected features ( Figures , S8, and S9 ). On the training data set with all features considered, the seven algorithms demonstrated strong classification performance, with the four assessed metrics ranging from about 75% to more than 96% ( Figure S8 ). Against the testing data set, we observed that k -nearest neighbors, gradient boosting, and logistic regression demonstrated the best classification performance, where k -nearest neighbors achieved high accuracy and precision of above 95%, while those of gradient boosting and logistic regression were above 90% ( Figure a). In addition, as the best-performing algorithm, the k -nearest neighbors misclassified only one cancerous tumor ( Figure b), and it had an AUC value of 95% ( Figure c). Gradient boosting was the next best-performing algorithm, and it correctly classified all cancerous tumors, but misclassified two benign tumors as cancerous. Additionally, the algorithm achieved an AUC value of 97.8%. Supervised classification of ovarian tumors in postmenopausal patients. (a) Table summarizing the performance of supervised machine learning algorithms on testing data set taking all features into account. (b) Confusion matrices and (c) ROC curves of the algorithms with better performance (i.e., k -nearest neighbors and gradient boosting). (d) Table summarizing the performance of supervised machine learning algorithms on testing data set taking only the most essential features into account. (e) Confusion matrices and (f) ROC curves of the algorithms with better performance (i.e., k -nearest neighbors and neural network). Finally, we assessed the algorithm performance on the data set of postmenopausal patients containing only the most essential features. We noted that all algorithms achieved more than 90% classification accuracy, precision, recall, and F 1 score on the training data set ( Figure S9 ). Against the testing data set, k -nearest neighbors and neural network emerged as the top performers, boasting accuracy and precision of 90.5 and 92.9%, respectively ( Figure d). Their confusion matrices revealed that both algorithms misclassified two cancerous tumors ( Figure e). Evaluation of the ROC curves showed that these two algorithms achieved an AUC value of 97.8% ( Figure f). It is noteworthy that, in our study, the 10-fold cross-validation was performed on the training data sets, while feature scoring and ranking were conducted using the complete data sets. This may raise the concern that data leakage might have been introduced during the classifier training process, which might inflate the performance metrics of the classifiers. To assess this possibility, we sought to perform feature scoring and ranking based on information gain ratio using only the training data sets ( Figures S10–S12 ). We noted that for the training data set with all patients, HE4, menopausal status, CA125, age, albumin, alkaline phosphatase, lymphocyte ratio, and platelet count were the most important features ( Figure S10 ). This corroborated our earlier finding using the complete data set, where the same eight features ranked the highest. Similarly, the most essential features for the premenopausal patient training data set, particularly HE4, albumin, CA125, red blood cell distribution width, age, lymphocyte ratio, globulin, and mean corpuscular hemoglobin ( Figure S11 ), also ranked the highest for the complete data set. Finally, a closely similar trend was also observed from the postmenopausal patient training data set ( Figure S12 ). All of these collectively suggest that there was minimal data leakage in our supervised machine learning analysis.

In this work, we capitalized on machine learning to analyze clinically relevant data on patient ovarian tumors. First, we sought to investigate the correlation between the types of ovarian tumors and the characteristic features of patients and to identify potential features that may be essential for differentiating between cancerous and benign tumors. Feature scoring and ranking unveiled that apart from the typical biomarkers associated with ovarian cancer, especially CA125 and HE4, the menopausal status of patients may be an essential feature that can be used to improve ovarian cancer detection. Specifically, a higher proportion of cancerous tumors could be found in postmenopausal patients than in premenopausal patients, suggesting that menopausal status may play a role in the incidence of ovarian cancer. Another feature that may also be useful for ovarian tumor classification is the concentration of albumin. In particular, the median concentrations of albumin for all patients were consistently lower in patients with cancerous tumors than in those with benign tumors. Interestingly, feature scoring and ranking also revealed that important attributes for distinguishing between cancerous and benign ovarian tumors may be unique to specific patient populations. For instance, although age ranked highly as a feature for distinguishing tumors in premenopausal patients, it may be less useful for classifying tumors in postmenopausal patients. Next, we were motivated to explore the feasibility of using supervised machine learning to classify ovarian tumors in different patient populations. We noted that the supervised machine learning algorithms examined in this study performed well during training, with the algorithms exhibiting classification accuracy, precision, recall, and F 1 score between about 75% and close to 98%. Analysis of the algorithm performance on various training data sets revealed that random forest and gradient boosting consistently emerged as some of the best algorithms for distinguishing the ovarian tumors of all and premenopausal patients, while k -nearest neighbors and gradient boosting showed the best performance on the postmenopausal patient data sets. An important highlight of our study is the analysis of the impact of feature selection and patient grouping on the classification performance of the supervised machine learning algorithms. Specifically, by analyzing only some of the most essential features, the predictive performance of nearly all algorithms on all, premenopausal, and postmenopausal patient data sets improved significantly. This highlights the significance of feature selection in enhancing the classification accuracy and precision of predictive algorithms. Comparing the algorithm performance across various patient population data sets, we noted that most algorithms displayed the best predictive performance on the postmenopausal patient data sets, underscoring the complexity of ovarian cancer diagnosis, which may vary depending on the patient demographics. It is noteworthy that the patient data set employed in our work was also used in several previously published reports. − Although the same data set was used, these studies were motivated by objectives different from ours. For example, most of the earlier works focused only on certain biomarkers, such as HE4 and CEA, and some established algorithms, including logistic regression and decision tree as well as k -nearest neighbors and support vector machine. The latter studies, however, started considering more supervised machine learning algorithms, including random forest, various gradient boosting algorithms, and stochastic gradient descent, , although these studies treated the patient data set differently prior to machine learning analysis. Most of the previously published reports also employed certain feature selection or feature weighting methods, such as minimum redundancy maximum relevance feature selection, adaptive differential evolution feature weighting, and the Shapley explainable method, in analyzing the patient data sets. As opposed to these studies, we adopted a distinct approach in data treatment, feature selection, and data analysis and identified possible ways to enhance the classification performance of supervised machine learning algorithms. We first preprocessed the patient data set by removing incomplete entries, ensuring that all patient data had complete and comparable features. Next, we performed feature selection based on information gain ratio and Gini index, followed by assessing the performance of seven supervised machine learning algorithms on patient data sets with all 47 features and with only 10 essential features. Simultaneously, we grouped the patients based on their menopausal status and analyzed the algorithm performance in classifying tumors in individual groups. Through this approach, we systematically evaluated the capabilities of different algorithms and determined optimal ways to improve the classification metrics of these algorithms. While our study has its merits, as emphasized in this section, we acknowledged that it also has several limitations. These limitations revolve around the characteristics of the data sets and the implemented analytical approaches. First, in terms of its characteristics, the data set used in this study is limited in data size, number of clinical features, and patient population diversity. After preprocessing, the data set consisted of a total of 309 entries, which comprised 140 patients with cancerous ovarian tumors and 169 patients with benign ovarian tumors. In addition, although our analysis of the menopausal status of patients offered valuable insights into the effect of demographic differences in ovarian cancer prediction, we did not manage to explore the impact of other demographic factors, such as ethnicity or socioeconomic status, which may also play a role in ovarian cancer incidence. , This is because all patient data used in this study were collected from a single hospital in one country, which significantly limited the scope of our work. It is noteworthy that the performance of supervised machine learning algorithms is generally influenced by both the quantity and quality of the employed data set, and any missing information or biases present in the data set may affect the results of the analysis. Also, the robustness of our supervised classifiers has not been assessed using either external data sets from different hospitals and countries or noisy and heterogeneous data sets from patients with other cancers or diseases. As such, the relatively small size of data set, low number of clinical features, and homogeneous patient populations examined in our work coupled with the lack of external validation data sets may affect the generalizability of our findings. Second, in terms of analytical approaches, although we employed widely used feature selection techniques to identify essential features for classifying ovarian tumors, the selection process itself might introduce bias. It is then possible that the selected features might not fully capture the complexity of the disease, potentially overlooking other important predictive factors. Next, despite our efforts in fine-tuning algorithm hyperparameters during the training process, there was a risk that some algorithms might memorize the noise in the training data sets rather than learning the true underlying patterns. This might lead to inflated performance metrics during training and reduce generalizability to unseen data sets. In parallel to exploring the use of machine learning in facilitating ovarian cancer diagnosis, it may also be worthwhile to consider the broader ethical and societal implications of such artificial-intelligence-based technologies. To start with, the potentially limited generalizability of our approach due to the use of a single-source data set with a relatively homogeneous patient population may introduce bias when the same approach is applied to more diverse patient populations. Next, the black box nature of some of the algorithms explored in this work may impede the wider deployment of machine-learning-based techniques, especially in clinical settings. Therefore, moving forward, in addition to technical performance, any future work in this area should prioritize the use of larger and more heterogeneous internal and external data sets coupled with more interpretable and transparent models to improve the inclusivity, equity, and accountability of machine learning applications in healthcare. Overall, despite its limitations, our study is highly significant. We believe that this work will provide deeper insight into the use of machine learning and relevant computational tools to advance ovarian cancer diagnosis and management to improve patient outcomes.

Ovarian cancer is one of the most lethal gynecological cancers with a five-year survival rate below 50%. Most of the ovarian cancer cases are typically diagnosed at later stages when the primary tumors have spread beyond the peritoneal cavity. − If detected earlier when the tumors are still confined within the ovaries, the five-year survival rate of ovarian cancer patients can be higher than 90%. As a complex disease, the onset of ovarian cancer is influenced by various factors, including age and menopausal status. − As women age and experience menopausal transition, their ovaries may undergo physiological changes, which may then influence the production of essential hormones such as estrogen. This hormonal shift has been reported to contribute to the pathogenesis of ovarian cancer. , Current ovarian cancer diagnostic methods include pelvic examination, imaging tests like transvaginal ultrasound (TVUS) and magnetic resonance imaging (MRI), and blood tests detecting the presence of common ovarian cancer biomarkers like cancer antigen 125 (CA125), CA19–9, carcinoembryonic antigen (CEA), human epididymis protein 4 (HE4), and α fetoprotein (AFP). − Pelvic examination, as a routine part of gynecological check-ups, is used to detect ovarian tumors, but this method lacks specificity in differentiating between benign and cancerous tumors. , Nonspecific findings such as pelvic masses can also be due to other conditions like cysts or fibroids. Furthermore, pelvic examination may miss early-stage tumors due to its limited sensitivity. While TVUS allows for the visualization of the ovaries and can identify abnormalities such as cysts or masses, it cannot definitively determine whether these abnormalities are benign or cancerous. , Similarly, MRI can provide detailed images of the ovaries and the surrounding tissues, but it may not be able to reliably assess the malignancy of tumors. Crucially, some of these methods may be invasive, which may cause discomfort in patients during examinations. As a less invasive alternative, blood tests detecting cancer biomarkers are attractive for several reasons. First, blood tests may potentially facilitate earlier ovarian cancer detection. − The simplicity and convenience of blood tests render them accessible to a wider population, which may enhance screening efforts. Second, blood tests can enable more frequent and regular monitoring of cancer progression or recurrence, allowing for more timely intervention when necessary. Finally, blood tests are more cost-effective than more complex imaging-based procedures. While a large number of biomarkers can be analyzed through blood tests, including blood constituents (e.g., erythrocyte count, leukocyte count, neutrophil count, and platelet count), electrolytes (e.g., potassium, sodium, chloride, and calcium), and cancer-related proteins (e.g., CA125, CA19–9, CEA, HE4, and AFP), and some of these may be indicative of the onset of ovarian cancer, − not all are essential for detecting this cancer. Furthermore, even for the common ovarian cancer biomarkers, their expression and concentration can be influenced by factors other than the presence of cancer. This may then lead to false positives or false negatives. For example, although CA125 is widely used as a biomarker for ovarian cancer, the expression of this protein can also be elevated in noncancerous conditions, particularly endometriosis and pelvic inflammatory diseases. , − Other biomarkers like CA19–9, CEA, HE4, and AFP may also lack specificity as their concentrations can be elevated in both benign and malignant conditions. , Consequently, there is a need to bridge this major gap in ovarian cancer diagnosis to improve the accuracy and reliability of ovarian cancer detection. In recent years, increasingly advanced computational methods, particularly machine-learning-based approaches, have been developed and exploited for various biomedical and healthcare applications, − including cancer diagnosis. − With their capacity to process large data sets, identify previously hidden patterns within complex data sets, and improve their predictive performance iteratively, machine learning models may be capitalized on to distinguish between benign and cancerous tumors. − Because of these attributes, there have been active explorations into the use of machine learning to aid the detection of various types of cancer, including breast, cervical, and ovarian cancers. − Herein, we employed machine-learning-based approaches to analyze clinically relevant data on patient ovarian tumors to identify potential clinical biomarkers that may correlate better with ovarian cancer. Specifically, we examined data sets comprising data on more than 300 female patients with 47 features, including the age and menopausal status of patients, the count and ratios of different blood cells, and the concentrations of various electrolytes, enzymes, and common ovarian cancer biomarkers. Based on feature scoring and ranking, we delineated several features that may be more useful than the rest in distinguishing cancerous tumors from benign ones in different female patient groups. We then assessed the classification performance of numerous supervised machine learning algorithms and identified effective combinations of algorithms and features to enhance ovarian cancer prediction. We anticipate that this study will provide deeper insight into the use of machine learning to improve the accuracy and reliability of blood tests for ovarian cancer diagnosis.