Machine Learning-Based Cervical Cancer Risk Prediction System: A Data-Driven Early Screening Model

This is an open-source machine learning project aimed at providing AI-assisted decision support for early cervical cancer screening through data preprocessing, exploratory data analysis, feature engineering, and classification model construction. Maintained by marwa189, the source code is available on GitHub (link: https://github.com/marwa189/cervical-cancer-risk-prediction) and was released on May 31, 2026. Its core objectives include: gaining an in-depth understanding of the distribution and correlations of cervical cancer risk factors, extracting effective predictive features, training and comparing multiple classification models, outputting personalized risk scores, and providing interpretability analysis of model decisions.

Cervical cancer is the fourth most common cancer among women globally, with over 600,000 new cases and more than 340,000 deaths each year. HPV infection is the main cause, and it takes 10-20 years from infection to canceration, providing a time window for early intervention. Traditional screening relies on Pap smears and HPV tests, which are effective but have issues such as high cost, strong dependence on medical resources, and insufficient screening coverage. Especially in resource-poor areas, the proportion of late diagnosis is high and the prognosis is poor. Machine learning technology can identify high-risk groups by analyzing data such as demographic characteristics, lifestyle, and medical history, optimize screening strategies, and provide new possibilities for solving the above problems.

The project uses a public dataset of cervical cancer risk factors, recording women's health information and diagnosis results. Feature categories include:

• Demographics: Age, marital status, education level, income level
• Lifestyle: Smoking history, drinking history, dietary habits, exercise status
• Medical history: Previous gynecological disease history, hormone use history, contraceptive use history, previous screening history
• Sexual behavior-related: Age at first sexual intercourse, number of sexual partners, HPV infection history, sexually transmitted disease history The target variable is the cervical cancer diagnosis result (binary classification: positive/negative).

The project's technical workflow covers the entire process:

1. Data Preprocessing: Handle missing values (delete features with high missing rates, fill with median/mode, etc.), detect outliers (IQR/Z-score, box plots), convert data types (encode categorical variables, standardize numerical variables).
2. Exploratory Data Analysis (EDA): Univariate analysis (distribution histograms, frequency distributions), bivariate analysis (relationship between features and target, chi-square test/T-test), multivariate analysis (correlation heatmap, PCA dimensionality reduction) to identify significantly correlated risk factors.
3. Feature Engineering: Select features via filter methods, wrapper methods, and embedding methods; construct combined/binned/ratio features; perform standardization/normalization/log transformation.
4. Model Construction: Implement multiple classification algorithms, including logistic regression (baseline model), decision trees, random forests, SVM, gradient boosting trees (XGBoost/LightGBM), and neural networks (MLP).

Model Evaluation: Use training/validation/test set splitting and K-fold cross-validation (stratified sampling); evaluation metrics include accuracy, precision, recall (more important in medical scenarios to avoid missed diagnoses), specificity, F1 score, AUC-ROC, and AUC-PR; analyze confusion matrices (focus on false negatives). Hyperparameter optimization uses grid search or random search. Interpretability: Provide feature importance (tree models, permutation importance, SHAP values), individual prediction explanations (LIME, decision path visualization), and rule extraction to ensure doctors and patients understand the basis of predictions.

Application Scenarios:

• Risk-stratified screening: Classify into high/medium/low risk based on risk scores to optimize screening intervals and resource allocation;
• Resource optimization: Prioritize services for high-risk groups to improve screening coverage;
• Health education: Identify high-risk behaviors and develop targeted strategies;
• Clinical research: Support epidemiological research and intervention effect evaluation. Limitations: Data may have biases (regional/population), self-report accuracy issues; model generalization ability needs verification; there are risks of privacy protection, psychological impact, and discrimination; integration with existing clinical processes, gaining doctor trust, and regulatory approval are required.

Future Directions:

• Multimodal data fusion: Integrate genomics, imaging, laboratory tests, and electronic health records;
• Deep learning applications: Automatic feature extraction and handling high-dimensional data;
• Real-time prediction system: Develop web/mobile applications to provide real-time risk assessment;
• Causal inference: Shift from correlation to identifying causal risk factors to support intervention strategies. Conclusion: This project demonstrates the potential of ML in cervical cancer risk prediction and provides technical support for early screening. It should be emphasized that ML models are auxiliary tools and cannot replace professional diagnosis. Ethics, privacy, and fairness need to be considered in applications. With technological progress, AI-assisted screening will become more accurate and widespread, contributing to global women's health.