The rapid adoption of machine learning–based classification systems in safety-critical and data-intensive domains has exposed the limitations of relying solely on accuracy as a primary evaluation criterion. In real-world scenarios characterized by class imbalance, asymmetric misclassification costs, and evolving data distributions, accuracy often provides an incomplete or misleading representation of model performance. This study presents a systematic review of widely used classification evaluation metrics—namely Precision, Recall, F1-score, and Receiver Operating Characteristic–Area Under the Curve (ROC-AUC)—to establish a more comprehensive framework for assessing predictive reliability and operational suitability. A structured review methodology was employed following established evidence synthesis protocols, examining peer-reviewed studies published between 2015 and 2025 across major digital libraries. The analysis incorporates empirical findings derived from benchmark datasets such as medical diagnostic repositories, financial fraud detection corpora, and network intrusion datasets (e.g., UNSW-NB15 and CICIDS), where classification decisions directly influence risk management and resource allocation. The review further evaluates the behavior of these metrics across diverse algorithmic paradigms, including logistic regression, support vector machines, random forests, gradient boosting models, and deep neural networks, under varying conditions of class imbalance, threshold selection, and cost sensitivity. The findings demonstrate that metric selection significantly affects model interpretation, deployment decisions, and system accountability, particularly in high-stakes environments such as healthcare diagnostics, cybersecurity threat detection, and financial anomaly monitoring. This work contributes a consolidated comparative analysis of evaluation metrics and offers practical, context-aware recommendations to guide researchers and practitioners in selecting appropriate performance measures for robust and transparent classification model assessment.

Classification Evaluation Metrics, Precision and Recall, F1-Score, ROC Curve, Imbalanced Datasets, Model Evaluation, Machine Learning Performance