Fuzzy-rank feature selection plus H2O AutoML ensembles reach up to 95.1% accuracy and 98.1% AUC on two public cervical cancer datasets

It proposes the Fuzzy Rank Feature Selection (FRFS) algorithm, which evaluates and selects the best predictor set along four dimensions: statistical significance, information gain, clinical importance, and uncertainty. Compared with methods relying on traditional feature selection techniques, this work folds uncertainty into the feature ranking criterion so that selection reflects both statistical and clinical considerations. The text describes this as an algorithmic procedure and states that the selected features are then used by H2O AutoML; the abstract does not give per-dimension weights or ablation results.

The features chosen by FRFS feed into H2O AutoML, which automatically builds ensemble models, and the framework is evaluated independently on two public cervical cancer datasets. Compared with pipelines that depend on manual optimization, model construction is delegated to AutoML, reducing manual tuning steps. Evaluation uses stratified five-fold cross-validation, and SMOTE is applied only to the training set in each fold to avoid information leakage, which strengthens the credibility of the reported results.

SHAP provides global feature importance and LIME provides patient-level explanations, giving interpretability at two levels. Compared with black-box predictive models that lack interpretability, the framework produces both overall and individual-level explanatory output. The abstract states that SHAP and LIME are used, but it does not show the specific explanation outputs or any clinician-facing readability assessment.

On the two public datasets the framework outperforms traditional machine learning models and the baseline H2O AutoML framework, with a highest accuracy of 95.1% and AUC of 98.1%. Relative to the baseline H2O AutoML and traditional machine learning models, adding uncertainty-aware feature selection and two-level explainability is reported to yield higher performance metrics. Evidence comes from independent experiments on two public datasets with stratified five-fold cross-validation; the abstract does not list the comparison models, confidence intervals, or statistical tests.

The framework targets the specific task of cervical cancer risk prediction and fits settings with structured tabular features where automated modeling and interpretable output are both wanted; its design goal is to let researchers and clinical stakeholders obtain accurate, robust, and interpretable predictions without manual tuning, and to understand model reasoning through SHAP global importance and LIME patient-level explanations. Because evaluation rests on two public datasets, the results apply to populations and feature systems resembling those data.

The available text is an abstract-level description: it contains no figures, no list of comparison models, no per-dataset breakdown, no weights for the four FRFS dimensions, and no ablation study, nor does it say whether the SHAP and LIME explanations were assessed by clinical staff. A careful reader would therefore still watch whether performance is consistent across the two datasets, how large the gain over baselines is, whether the selected features are clinically interpretable, and how the pipeline behaves on external data or other populations. These are open questions for follow-up validation.