A Systematic Review Of Explainable Artificial Intelligence Techniques For Trustworthy Machine Learning Systems

While machine learning models become increasingly predictive, their lack of transparency threatens trust in high-risk domains like healthcare, finance, and civil infrastructure. Explainable AI research, thus, mainly deals with the challenges associated with making model behaviors and decision processes interpretable. This systematic review, carried out using the PRISMA 2020 statement, examines 89 peer-reviewed Q1 and Q2 journal articles published from 2018 to 2025 and identifies fourteen different XAI techniques. The leading methods in the literature are post-hoc explainability (82%), while SHAP and LIME are the most widely adopted XAI techniques, more so in healthcare applications at 28%. Other model-specific techniques include the Grad-CAM method and attention mechanisms, which find wide applications in computer vision and natural language processing tasks. Going beyond descriptive syntheses, this review proposes an integrated hybrid framework for explainability that leverages SHAP with counterfactual explanations, enhancing interpretive, actionable, and user trust. The review further develops key gaps in current research inquiries: (i) absence of causal reasoning mechanisms, (ii) lacks of uniform evaluation metrics, and (iii) limited human-centered validation. Directions for further studies are discussed and should be oriented toward understanding causal XAI, federated and privacy-preserving explainability, and neurosymbolic hybrid models.