What question did this study set out to answer?

The aim is to enhance cervical cancer risk stratification using an MLOps framework that leverages routine clinical data.

April 21, 2026Open Access

CervixNet-Ensemble: A Stacked Meta-Learning MLOps Framework for Robust Biopsy-Confirmed Cervical Cancer Risk Stratification Leveraging Multi-Modal Behavioral and Clinical Predictors

Key Points

The aim is to enhance cervical cancer risk stratification using an MLOps framework that leverages routine clinical data.
Developed a multi-model machine learning pipeline combining several algorithms including XGBoost and Random Forest.
Utilized data preprocessing techniques like SMOTE and feature reduction with mutual information.
Implemented the model on the UCI Cervical Cancer Risk Factors Dataset and validated results using cross-validation.
Achieved a biopsy-confirmed sensitivity of 97.37% and specificity of 99.38%.
Overall accuracy reached 99.42%, with an AUC-ROC of 0.989.
The framework outperformed existing models and thresholds on the given dataset.

Abstract

Cervical cancer causes devastating deaths greater than 340000 annually worldwide, but it is one of the most identifiable and preemptable ailments under the condition that suitable screening devices are available. The need to use costly laboratory tests and interpretation by specialists significantly impedes clinical risk stratification in low-resource setting. The article presents CervixNet-Ensemble, a clinically oriented, end-to-end machine learning applications (MLOps) pipeline that operationalizes risk classification with biopsy confirmation using merely structured, routinely gathered behavioral and demographic information. The framework trains four gradient-boosting/bagging base learners, XGBoost, LightGBM, CatBoost, and Random Forest, and trains a logistic regression meta-learner to be trained on cross-validated cross-fold probability predictions with the base learners, thus learning a principled, data-guided combination of heterogeneous classifier inductive biases. On the publicly available UCI Cervical Cancer Risk Factors Dataset (n = 858), the system uses an inspired preprocessing pipeline which includes K-Nearest neighbour imputation, Synthetic Minority Oversampling Technique (SMOTE) and mutual information guiding feature reduction. Maximal choice of decision boundaries using the Youden J Statistic values provides a biopsy-confirmed sensitivity of 97.37, specificity of 99.38, overall accuracy of 99.42 and AUC-ROC of 0.989 which is better than any single model choice threshold or published threshold to date on this data set. The full system is implemented as a live, publicly available Streamlit clinical decision support application with SHapley Additive exPlanations (SHAP) to attribute features transparently and per-patient and automatically produce PDF diagnostic reports. CervixNet-Ensemble illustrates that clinical data of near-perfect interpretable and instantaneously deployable cervical cancer risk stratification can be obtained using routinely available clinical information with direct relevance to screening based healthcare settings across the globe.

CervixNet-Ensemble: A Stacked Meta-Learning MLOps Framework for Robust Biopsy-Confirmed Cervical Cancer Risk Stratification Leveraging Multi-Modal Behavioral and Clinical Predictors

Key Points

Abstract

Cite This Study