Detecting human motion and predicting human intentions by analyzing body signals are challenging but fundamental steps for the implementation of applications presenting human–robot interaction in different contexts, such as robotic rehabilitation in clinical environments, or collaborative robots in industrial fields. Machine learning techniques (MLT) can face the limit of small data amounts, typical of this kind of applications. This paper studies the illustrative case of the reaching movement in 10 healthy subjects and 21 post-stroke patients, comparing the performance of linear discriminant analysis (LDA) and random forest (RF) in: (i) predicting the subject’s intention of moving towards a specific direction among a set of possible choices, (ii) detecting if the subject is moving according to a healthy or pathological pattern, and in the case of discriminating the damage location (left or right hemisphere). Data were captured with wearable electromagnetic sensors, and a sub-section of the acquired signals was required for the analyses. The possibility of detecting with which arm (left or right hand) the motion was performed, and the sensitivity of the MLT to variations in the length of the signal sub-section were also evaluated. LDA and RF prediction accuracies were compared: Accuracy improves when only healthy subjects or longer signals portions are considered up to 11% and at least 10%, respectively. RF reveals better estimation performance both as intention predictor (on average 59.91% versus the 62.19% of LDA), and health condition detector (over 90% in all the tests).