Cervical cancer prediction using machine learning models based on blood routine analysis

Abstract

Background and objective: Cervical cancer is the fourth most common cancer among women globally. The key of prevention and treatment of cervical cancer is early detection, diagnosis and treatment. We aimed to develop an interpretable model to predict the risk for patients with cervical cancer based on blood routine data and used the Shapley additive interpretation (SHAP) method to explain the model and explore factors for cervical cancer. Methods In this paper, medical records of patients from 2013 to 2023 were collected for retrospective study. 2533 patients with cervical cancer were used as the case group, and 9879 patients with apparent healthy subjects were used as the control group. Using age, clinical diagnosis information and 22 blood cell analysis results, four different algorithm were used to construct cervical cancer prediction model. Results Using lasso regression and random forest method, 15 important blood routine features were finally selected from 23 features for model training. Comparatively, the XGBoost model had the highest predictive performance among four models with an area under the curve (AUC) of 0.964, whereas RF had the poorest generalization ability (AUC = 0.907). The SHAP method reveals the top 6 predictors of cervical cancer according to the importance ranking, and the average of the PDW was recognized as the most important predictor variable. Conclusion In conclusion, we select the best ML based on performance and rank the importance of features according to Shapley Additive Explanation (SHAP) values. Compared to the other 4 algorithms, the results showed that the XGB had the best prediction performance for successfully predicting cervical cancer recurrence and was adopted in the establishment of the prediction model.