Tele-assessment systems are crucial for home-based rehabilitation, as they allow therapists to assess the status of patients and adjust the parameters of various home-based training devices. Traditional force/torque sensors are commonly used in tele-assessment systems to detect muscle strength because such sensors are convenient. However, muscle activity can be measured using surface electromyography (sEMG), which records the activation level of skeleton muscles and is a more accurate method for determining the amount of force exerted. Thus, in this paper, a method for predicting muscle strength using only sEMG signals is proposed. The sEMG signals measure the isometric downward touch motions and are recorded from four muscles of the forearm. The prediction function is derived from a musculoskeletal model. The parameters involved are calibrated using the Bayesian linear regression algorithm. To avoid the complex modeling of the entire movement, a neural network classifier is trained to recognize the force-exerting motion. Experimental results show that the mean root-mean-square error of the proposed method is below 2.5 N. In addition, the effects of the high-pass cutoff frequency and the co-activation of flexors and extensors for EMG force prediction are discussed in this paper. The performance of the proposed method is validated further in real-time by a remote predicted-force evaluation experiment. A haptic device (Phantom Premium) is used to represent the predicted force at the therapist’s remote site. Experimental results show that the proposed method can provide acceptable prediction results for tele-assessment systems.