<p>Machine learning (ML) techniques are increasingly applied to support software defect prediction. The effectiveness of ML models in this domain is highly dependent on the appropriate choice of hyperparameters, feature subsets, and optimization objectives. To address these challenges, a growing body of research investigates the use of metaheuristic algorithms to optimize ML models for software defect prediction. Despite this growth, a comprehensive and critical synthesis of existing empirical evidence remains lacking. This study presents a systematic literature review and critical analysis of research applying metaheuristic optimization to ML models in software testing for defect prediction task. A rigorously defined review protocol is employed to identify primary studies published in major software engineering venues. The review analyzes the types of ML models and metaheuristic algorithms used, the software defect prediction addressed, the optimization goals and evaluation metrics considered, and the datasets adopted for training and empirical validation. Unlike prior surveys that primarily provide descriptive overviews or focus on either defect prediction or metaheuristic optimization in isolation, this study emphasizes a critical assessment of empirical practices, with a particular focus on dataset usage, evaluation rigor, and reproducibility-related reporting. Additionally, by restricting the analysis to recent journal publications (2021–2025), the review captures current methodological trends and reflects the maturity of empirical practices in the field, enabling a more accurate interpretation of contemporary research directions and limitations. In addition to descriptive synthesis, a critical assessment of empirical practices is provided. The analysis reveals recurring limitations, including heavy reliance on a small number of publicly available datasets, insufficient baseline comparisons, limited use of statistical significance testing, and inadequate reporting for reproducibility. These findings raise concerns regarding the applicability to modern and industrial software systems and robustness of reported results. Based on the identified trends and shortcomings, key research gaps and future directions are outlined to support more rigorous, reproducible, and empirically sound research on metaheuristic-optimized ML for software defect prediction.</p>