Exploring Ovarian Cancer Prediction Models and Potential Markers Using Machine Learning.

OBJECTIVE

To develop machine learning models, facilitate a more accurate diagnosis of ovarian cancer (OC), and explore potential markers.

METHODS

Overall, 311 patients diagnosed with OC, 56 with borderline ovarian tumors (OTs), and 368 patients with benign OTs were defined as derivation cohort and randomly divided into training (70%) and internal validation (30%) sets. An independent external validation cohort was also established. A total of 34 variables including patients' demographic characteristics and laboratory test results were collected. Models were developed using artificial neural network, support vector machine, random forest, and extreme gradient boosting (XGBoost).

RESULTS

All four models achieved high accuracy, with XGBoost achieving the highest area under the curve (AUC). When using the XGBoost model to differentiate OC from borderline and benign OTs, the AUC and 95% confidence interval, sensitivity, specificity, positive predictive value, and negative predictive value of the training set were 0.973 (0.962-0.985), 84.2%, 96.6%, 93.9%, and 90.6%, respectively. For the internal validation set, the values were 0.932 (0.897-0.966), 74.7%, 92.0%, 85.5%, and 85.2%. The eight most important variables were human epididymis protein 4, carbohydrate antigen 125, lactate dehydrogenase, D-dimer, age, testosterone, follicle-stimulating hormone, and hemoglobin. Subgroup analyses also revealed that this model exhibited outstanding performance in identifying early-stage OC and epithelial OC.

CONCLUSION

Machine learning models demonstrate excellent accuracy in distinguishing OC from borderline and benign OTs, with several potential markers being validated.