The purpose is to accurately identify women at high risk of developing cervical cancer so as to optimize cervical screening strategies and make better use of medical resources. However, the predictive models currently in use require clinical physiological and biochemical indicators, resulting in a smaller scope of application. Stacking-integrated machine learning (SIML) is an advanced machine learning technique that combined multiple learning algorithms to improve predictive performance. This study aimed to develop a stacking-integrated model that can be used to identify women at high risk of developing cervical cancer based on their demographic, behavioral, and historical clinical factors.
The data of 858 women screened for cervical cancer at a Venezuelan Hospital were used to develop the SIML algorithm. The screening data were randomly split into training data (80%) that were used to develop the algorithm and testing data (20%) that were used to validate the accuracy of the algorithms. The random forest (RF) model and univariate logistic regression were used to identify predictive features for developing cervical cancer. Twelve well-known ML algorithms were selected, and their performances in predicting cervical cancer were compared. A correlation coefficient matrix was used to cluster the models based on their performance. The SIML was then developed using the best-performing techniques. The sensitivity, specificity, and area under the curve (AUC) of all models were calculated.
The RF model identified 18 features predictive of developing cervical cancer. The use of hormonal contraceptives was considered as the most important risk factor, followed by the number of pregnancies, years of smoking, and the number of sexual partners. The SIML algorithm had the best overall performance when compared with other methods and reached an AUC, sensitivity, and specificity of 0.877, 81.8%, and 81.9%, respectively.
This study shows that SIML can be used to accurately identify women at high risk of developing cervical cancer. This model could be used to personalize the screening program by optimizing the screening interval and care plan in high- and low-risk patients based on their demographics, behavioral patterns, and clinical data.