TL;DR
This paper introduces a novel cervical cancer prediction model (CCPM) that leverages data pre-processing techniques like outlier detection (DBSCAN, iForest) and data balancing (SMOTE, SMOTETomek) to improve the accuracy of Random Forest classification. The CCPM demonstrates superior performance compared to previous methods and has been integrated into a mobile application for early risk assessment.
What Problem Does This Paper Address?
Cervical cancer remains a significant global health concern, particularly in developing nations, due to its insidious early stages and the difficulty in distinguishing risk factors. Existing machine learning approaches for cervical cancer prediction often struggle with imbalanced datasets (where one class, e.g., healthy individuals, vastly outnumbers the other, e.g., those with cancer) and the presence of outliers, which can skew model training and reduce predictive accuracy. This paper aims to address these challenges by developing a more robust and accurate prediction model that effectively handles data imperfections.
Methods
The core of the proposed Cervical Cancer Prediction Model (CCPM) involves a multi-stage approach:
- Feature Extraction: Chi-square feature selection was employed to identify the most relevant risk factors from the dataset.
- Outlier Detection: Two distinct methods were used to identify and remove anomalous data points:
- DBSCAN (Density-Based Spatial Clustering of Applications with Noise): A clustering algorithm that groups together points that are closely packed together, marking points that lie alone in low-density regions as outliers.
- iForest (Isolation Forest): An ensemble learning method that isolates anomalies by building random trees, where outliers are expected to have shorter path lengths.
- Data Balancing: To address the issue of imbalanced datasets, two oversampling techniques were applied:
- SMOTE (Synthetic Minority Over-sampling Technique): Generates synthetic minority class samples by interpolating between existing minority class instances and their nearest neighbors.
- SMOTETomek: Combines SMOTE with Tomek links, which are pairs of nearest neighbors from different classes that are close to each other. Tomek links are then removed to clean up the data and further balance the classes.
- Classification: The Random Forest (RF) algorithm was chosen as the final classifier due to its robustness and strong performance in previous studies.
The CCPM was evaluated across four scenarios, combining each outlier detection method with each data balancing technique (DBSCAN + SMOTE, DBSCAN + SMOTETomek, iForest + SMOTE, iForest + SMOTETomek), all using Random Forest as the classifier.
Datasets
The study utilized a publicly available dataset from the UCI Machine Learning Repository, specifically the "Cervical Cancer (Risk Factors)" dataset. This dataset contains 858 instances with 36 features, collected from Hospital Universitario de Caracas in Venezuela. The paper does not introduce new benchmarks but rather uses this existing dataset for evaluation.
Key Insights and Ideas
- Data Pre-processing is Crucial: The study strongly emphasizes that effective outlier detection and data balancing are not just supplementary steps but are critical for achieving high accuracy in medical prediction tasks, especially with imbalanced datasets.
- iForest Outperforms DBSCAN: For outlier detection in this specific cervical cancer dataset, iForest proved to be more effective than DBSCAN, identifying a significantly larger number of outliers (86 vs. 2).
- SMOTETomek Shows an Edge: While both SMOTE and SMOTETomek improved data balance, SMOTETomek generally led to better model performance, suggesting its utility in cleaning noisy data introduced by SMOTE.
- Random Forest is a Strong Classifier: The study reaffirms Random Forest's capability as a powerful classifier for complex medical datasets, outperforming other popular machine learning algorithms like SVM, MLP, Logistic Regression, Naïve Bayes, and KNN.
- Integrated Approach Yields Superior Results: The combination of advanced pre-processing techniques (outlier detection and data balancing) with a robust classifier (Random Forest) resulted in a prediction model that significantly outperformed previous methods.
Key Results
The proposed CCPM, particularly the combinations involving iForest and SMOTETomek, achieved state-of-the-art performance. For instance, for the "Biopsy" target variable, the iForest + SMOTETomek + RF combination achieved an accuracy of 98.919%, and iForest + SMOTE + RF achieved 98.925%. These results were consistently higher than those obtained with DBSCAN-based approaches and significantly surpassed previous studies. The paper also highlights that the CCPM achieved better sensitivity, specificity, and accuracy compared to prior works by Wu and Zhu [10] and Abdoh et al. [22].
Furthermore, the practical implications of the CCPM were demonstrated through the development of a mobile application. This app allows users to input their risk factors, and the CCPM provides an early prediction of cervical cancer risk, aiming to facilitate timely intervention.
Limitations and Future Work
- Single Dataset Limitation: The study acknowledges that it was conducted using only one dataset. While the results are promising, applying the CCPM to diverse cervical cancer datasets would provide more robust validation.
- Computational Cost: The paper notes that the proposed CCPM, due to its combination of outlier detection, data balancing, and Random Forest, can be slower and require more memory compared to simpler algorithms. However, this trade-off is deemed acceptable given the improved accuracy.
- Future Work Suggestions:
- Explore more diverse outlier detection and over-sampling techniques.
- Apply each combination of techniques to CCPM to further optimize diagnosis performance.
- Extend the CCPM to diverse cancer datasets (e.g., breast, liver, lung, prostate, thyroid, kidney) to enhance clarity and quality of results across different cancer types.
Summary updated