A Comparative Study on Machine Learning Classifiers for Cervical Cancer Prediction: A Predictive Analytic Approach

Khandaker Mohammad Mohi Uddin, Iftikhar U. Sikder, Md. Nahid HasanEAI Endorsed Transactions on Internet of Things 2024 · DOI

TL;DR

This paper explores various machine learning classifiers for predicting cervical cancer risk, finding that hyperparameter-tuned Support Vector Machines (SVM) achieve a remarkable 99.64% accuracy. The research also introduces a web application to make these predictions accessible, aiming to improve early diagnosis and patient outcomes, especially in underserved regions.

What Problem Does This Paper Address?

Cervical cancer remains a significant global health challenge, particularly in low- and middle-income countries where access to preventive care and early detection is limited. Early identification of risk factors and timely diagnosis are crucial for effective treatment and improving patient outcomes. While machine learning (ML) has shown promise in medical diagnostics, a comprehensive comparison of various ML models for cervical cancer prediction, along with an accessible tool for its application, is needed. This study aims to bridge this gap by evaluating multiple ML algorithms and demonstrating their potential for accurate and early cervical cancer risk assessment.

Methods

The researchers employed a comparative approach, evaluating eleven different machine learning classifiers: Support Vector Machine (SVM), Naïve Bayes (NB), Decision Tree (DT), K-Nearest Neighbors (KNN), Random Forest (RF), Logistic Regression (LR), Gradient Boosting (GB), Nearest Centroid (NC), Multilayer Perceptron (MP), and AdaBoost (AB).

Key methodological steps included:

  • Data Preprocessing: This involved handling missing values (imputation with the mean), feature selection using a combination of SelectKBest and XGBoost to identify the most relevant 14 features, and addressing class imbalance using Random Over-sampling (ROS).
  • Model Training and Evaluation: The dataset was split into 80% for training and 20% for testing. Models were trained and then evaluated using standard metrics like accuracy, sensitivity, specificity, precision, and F-measure.
  • Hyperparameter Tuning: To optimize model performance, particularly for SVM, grid search was employed to fine-tune key hyperparameters (e.g., regularization parameter C, kernel coefficient γ, and kernel type).
  • Web Application Development: A user-friendly web application was developed using Flask to allow users to input their data and receive a risk prediction, making the ML model accessible.

Datasets

The study utilized a publicly available dataset from the Kaggle repository. This dataset comprises 835 patient samples with 36 characteristics, including demographic information, clinical history, and diagnostic test results. The dataset contains 54 cervical cancer cases and 781 non-cancerous cases, with the "Biopsy" column indicating the diagnosis (1 for cervical cancer, 0 for healthy).

Key Insights and Ideas

  • Hyperparameter Tuning is Crucial: The study strongly emphasizes that optimizing hyperparameters can significantly boost model performance. The SVM model, in particular, saw a dramatic improvement in accuracy after tuning, highlighting its importance beyond just selecting the right algorithm.
  • SVM's Superiority Post-Tuning: While Decision Trees initially showed high accuracy, the hyperparameter-tuned SVM ultimately achieved the best performance, demonstrating its robustness and adaptability for this specific prediction task.
  • Accessibility Matters: The development of a web application translates complex ML models into a practical tool for users, democratizing access to predictive healthcare and potentially improving early detection rates.
  • Feature Importance: The study identified key features like Schiller's assay, diagnosis history (Dx), and smoking habits as significant indicators for cervical cancer risk, aligning with known risk factors.

Key Results

  • Before Hyperparameter Tuning:
    • Decision Tree (DT) achieved the highest accuracy at 98.91%.
  • Random Forest (RF) followed closely with 98.54% accuracy.
  • After Hyperparameter Tuning:
    • Support Vector Machine (SVM) achieved a superior accuracy of 99.64%, with a precision of 99.26% and an F1-score of 0.9963.
  • The ROC-AUC curves indicated that SVM and Random Forest had the highest AUC values (1.00), signifying excellent discriminative power.
  • Web Application: A functional web application was developed, allowing users to input their data and receive a prediction result, indicating the probability of cervical cancer.

Limitations and Future Work

  • Dataset Size: While the dataset is publicly available, its size (835 samples) might be considered moderate for complex ML tasks. The authors suggest incorporating additional datasets, especially those with spiral drawings, to further enhance model capabilities.
  • Model Scalability: The authors aim to increase the model's scalability in future iterations.
  • Hybrid Architectures: Exploring hybrid architectures that combine ML with deep learning is proposed as a future development path to achieve more advanced and adaptable models.
  • Real-World Validation: Future work could involve validating the predictive model in real-world clinical settings and with diverse datasets to improve its robustness and applicability.
  • Feature Selection Refinement: While effective, further investigation into feature selection methods could potentially yield even better results.

Summary updated

Abstract

INTRODUCTION: Cervical cancer is a significant global health concern, particularly in underdeveloped nations where preventive healthcare measures are limited. Early identification of the risks associated with cervical cancer is essential for both prevention and treatment. OBJECTIVES: In recent years, machine-learning algorithms have gained popularity as potential techniques for determining a person's risk of developing cancer based on demographic and medical information. This study uses a dataset that contains patient demographics, clinical history, and results from diagnostic tests to examine how machine learning-based algorithms can be used to predict the risks of cervical cancer. METHODS: Various machine learning approaches are used to create predictive systems, including Support Vector Machine (SVM), Naïve Bayes (NB), Decision Tree (DT), K-Nearest Neighbors (KNN), Random Forest (RF), Logistic Regression (LR), Gradient Boosting (GB), Nearest Centroid (NC), Multilayer Perceptron(MP), and AdaBoost (AB). RESULTS: The prediction capability of these models is assessed using performance metrics such as accuracy, sensitivity, specificity, f-measure, precision, and area under the receiver operating characteristic curve (AUC-ROC). Our results show that the decision tree has the highest accuracy, precision, and f1-score (98.91%, 97.81%, and 0.9889). Additionally, model performance was optimized by the use of hyperparameter tuning. After hyperparameter adjustment, the Support Vector Machine (SVM) showed superior accuracy of 99.64%, precision of 99.26%, and an F1-score of 0.9963, thereby indicating its potential in cervical cancer probability prediction. We also created a web application that uses a machine-learning model to estimate the risk of cervical cancer. CONCLUSION: The findings of this study highlight the significance of SVM and demonstrate the potential and capabilities of machine learning techniques to enhance accurate prediction and patient outcomes for cervical cancer screening.
BibTeX
@article{uddin2024comparative,
  author = {Khandaker Mohammad Mohi Uddin and Iftikhar U. Sikder and Md. Nahid Hasan},
  title = {{A Comparative Study on Machine Learning Classifiers for Cervical Cancer Prediction: A Predictive Analytic Approach}},
  journal = {EAI Endorsed Transactions on Internet of Things},
  year = {2024},
  doi = {10.4108/eetiot.6223},
}
Figure