For understudied species like common vetch (Vicia sativa) and faba bean (Vicia faba), the combination of machine learning (ML) and meta-analysis (MA) has revolutionary promise for improving leguminous crops. Using MA and PRISMA-guided systematic review, this work synthesizes 115 peer-reviewed publications from 2015 to 2025 to assess machine learning applications in genomic and phenotypic trait prediction. The results show that ensemble approaches (e.g., Random Forest, XGBoost) perform better than standard models in disease resistance classification (AUC 0.88–0.91 via SVM) and yield prediction (R2 up to 0.92 in Phaseolus vulgaris). ML improves genomic selection (85–95% accuracy for flowering time GWAS) and root trait phenotyping (89% accuracy in faba bean drought adaptation) for Vicia species. Vicia villosa shows considerable phenotypic flexibility (CV 25–50%) but low model performance (F1-score 0.60–0.75 for winter survival), highlighting research gaps in tropical legumes, according to a meta-analysis. CNNs automate root architecture analysis (IoU 0.94); however, PLS regression is superior in NIRS-based nutritional trait prediction (R2 0.91 for protein). Data standards and the computing requirements for huge genomes (such the 13 Gb faba bean genome) are challenges. Precision breeding for nutritional quality and climatic resistance is made possible by the faster trait discovery made possible by the combination of ML and MA. In order to close the gap between model crops and ignored legumes, future efforts will focus on explainable AI, multi-omics integration, and cloud-based pipelines.