Stacking Machine Learning Algorithms for Biomarker-Based Preoperative Diagnosis of a Pelvic Mass

Simple Summary It is critical for women who are diagnosed with a pelvic mass, or an ovarian cyst to be accurately assessed for their risk of having an ovarian malignancy. Accurate risk stratification for these women will allow for appropriate triage and referral to centers best equipped to treat women diagnosed with ovarian cancer. In this study, machine learning (ML) algorithms were used to determine the optimal combination of biomarkers for prediction of malignancy in women presenting with a pelvic mass. Nine unique ML algorithms were employed to evaluate age, menopausal status, race, and levels of 67 biomarkers from serum, urine, and plasma samples prospectively collected in a cohort 140 women with a variety of pelvic mass diagnoses benign and malignant. A complex statistical algorithm using serum levels of CA125, HE4 and transferrin provided greater than 93% sensitivity and specificity for the preoperative prediction of malignancy in women presenting with a pelvic mass. Objective: To identify the most predictive parameters of ovarian malignancy and develop a machine learning (ML) based algorithm to preoperatively distinguish between a benign and malignant pelvic mass. Methods: Retrospective study of 70 predictive parameters collected from 140 women with a pelvic mass. The women were split into a 3:1 "training" to "testing" dataset. Feature selection was performed using Gini impurity through an embedded random forest model and principal component analysis. Nine unique ML classifiers were assessed across a variety of model-specific hyperparameters using 25 bootstrap resamples of the training data. Model predictions were then combined into an ensemble stack by LASSO regression. The final ensemble stack and individual classifiers were then applied to the testing dataset to assess model performance. Results: Feature selection identified HE4, CA125, and transferrin as three predictive parameters of malignancy. Assessment of the ensemble stack on the testing dataset outperformed all individual ML classifiers in predicting malignancy. The ensemble stack demonstrated an accuracy of 97.1%, a receiver operating characteristic (ROC) area under the curve (AUC) of 0.951, and a sensitivity of 93.3% with a specificity of 100%. Conclusions: Combining the measurement of three distinct biomarkers with the stacking of multiple ML classifiers into an ensemble can provide valuable preoperative diagnostic predictions for patients with a pelvic mass.