Poor pharmacokinetic behavior and toxicity remain major contributors to compound attrition in drug discovery, prompting rapid expansion of machine learning approaches for ADMET prediction as organizations seek earlier and more reliable risk assessment before costly experimental stages. This systematic review evaluates machine learning models for ADMET prediction published between 2017 and 2024, focusing on model types, endpoint coverage, molecular representations, validation practices, and evidence of translational readiness. A PRISMA 2020-compliant search strategy was applied to PubMed, Scopus, IEEE Xplore, and Web of Science, with records screened by two reviewers and eligible studies synthesized narratively according to ADMET endpoint and validation approach. The literature demonstrates a growing body of work across absorption, metabolism, and toxicity endpoints, with toxicity prediction and web-based ADMET platforms particularly prominent. Most studies relied on internal validation, while external, temporal, scaffold-based, and prospective validation were reported less consistently. Overall, machine learning for ADMET prediction has reached technical maturity in several endpoint areas, yet its translation into drug discovery practice remains limited by inconsistent validation standards, highlighting the need for better benchmarks, clearer applicability-domain reporting, and prospective evaluation.