Code Quality Alarms: A Review of Techniques, Datasets, and Emerging Trends in Detecting Smells and Anti-Patterns

This systematic literature review examines research on code smell and anti-pattern detection in software engineering from 2020 to April 2025, aiming to synthesise detection methodologies, evaluate dataset practices, and identify critical research gaps and trends. A total of 49 peer-reviewed studies were selected from five major scholarly databases: Google Scholar, Scopus, ScienceDirect, IEEE Xplore, and the ACM Digital Library. Approximately 80% of the studies utilised custom-built datasets derived from open-source repositories like GitHub, with limited adoption of public datasets such as Qualitas Corpus, Fontana, and MLCQ. This dataset fragmentation undermines reproducibility and limits cross-study comparability. To reveal methodological trends, the reviewed techniques were categorised into static analysis,  traditional machine learning, deep learning, and emerging paradigms such as large language models and human-in-the-loop approaches. This taxonomy highlights a shift toward hybrid frameworks that combine static analysis with AI-driven techniques, although standardised benchmarks remain scarce. Feature Envy was the most frequently studied code smell, underscoring its perceived impact on maintainability. Notably, 88% of studies focused exclusively on Java, revealing a strong language bias that restricts generalizability. The findings underscore the need for standardised, multilingual datasets, unified annotations, and broader adoption of advanced AI techniques to enhance scalability, practical relevance, and long-term software quality.