Cervical cancer remains one of the leading causes of cancer-related deaths among women globally, particularly in low- and middle-income countries. Early detection plays a crucial role in reducing the morbidity and mortality associated with cervical cancer. This study investigates the efficacy of advanced feature selection and machine learning techniques in the early detection of cervical cancer. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied, enhancing model performance on minority classes. A multilayer perceptron (MLP) classifier, guided by SHapley Additive exPlanations (SHAP), achieved a perfect classification accuracy using 14 selected features. This work also looks into the inter-relationship between the resampled data and analysis of the machine learning models for observing the predictive power. Among these, medical indicators such as Dx:HPV, Dx:CIN, and overall diagnostic features emerged as the most influential, a finding validated by both model interpretability metrics and Spearman’s correlation heatmaps, as well as supported by clinical literature. Additionally, a Stochastic Gradient Descent (SGD) model demonstrated exceptional performance with 99.39% accuracy, zero false negatives, and only three false positives, highlighting its high sensitivity and specificity. It also surpassed Explainable Boosting Machines (EBM) and Liquid Neural Networks (LNN) in reducing misclassification errors. These findings underscore the critical role of tailored feature engineering, robust model selection, and thorough data preprocessing in achieving clinically meaningful outcomes. The developed system can assist in enhancing the primary detection of cervical cancer and can be useful for health professionals in making diagnostic and therapeutic decisions.