<p>Low muscle mass is associated with adverse health outcomes not only in older adults, but also in young and middle-aged adults. However, traditional muscle mass assessments rely on specialized equipment and sufficient physical space, limiting their applicability in primary care and community settings. Additionally, machine learning-based approaches suffer from limited interpretability owing to their black-box nature. Therefore, we aimed to evaluate a machine learning-based approach using surface electromyography (sEMG) signals recorded during walking to classify muscle mass levels and identify key features underlying the model’s decisions in young and middle-aged adults. We enrolled 133 community-dwelling adults aged 20–59 years (71 men, 53.4%; 62 women, 46.6%). Appendicular skeletal muscle mass was measured using bioelectrical impedance analysis (InBody 970). Participants were divided into low- and high–muscle-mass groups using unsupervised k-means clustering. Five classifiers (logistic regression, support vector machine, k-nearest neighbors, random forest, and extreme gradient boosting) were trained, and performance was compared for normal and fast walking. Model interpretability was assessed with Shapley additive explanations. The low- and high- muscle-mass groups had mean ages of 36.12 and 39.84 years (<i>p</i> = 0.086), respectively. Extreme gradient boosting achieved the highest accuracy during normal walking (95%), whereas random forest performed best during fast walking (94%). During fast walking, a higher zero crossing rate of the biceps femoris contributed most to classifying higher muscle mass; during normal walking, a lower maximum power of the tibialis anterior was most influential. These findings demonstrate a feasible and interpretable approach to stratifying muscle mass from sEMG during gait. Larger and more diverse dataset are warranted to improve generalizability and support potential clinical application.</p>