Human musculoskeletal recognition is essential for analyzing the physical condition of elderly individuals. However, many elderly individuals are concerned about the inconvenience during rehabilitation and therapy sessions. To address this issue, we propose a surrogate model of the human musculoskeletal system with low computational cost, enabling its application on mobile devices. This paper focuses on validating the system for human upper limb parts. The proposed model employs a multi-layer perceptron architecture with a ReLU activation function. It consists of 5 input neurons, 128 neurons in the first hidden layer, 64 neurons in the second hidden layer, and 10 output neurons representing the activity of upper limb muscles. The input features include 3 joint angles of the glenohumeral joint, 1 joint angle of elbow flexion, and 1 load measurement in the hand. These inputs are obtained from a human skeleton recognition module and the camera on the mobile device. The dataset used for training is derived from human musculoskeletal simulations to mitigate the risks associated with experiments involving heavy loads. We conducted a training data comparison using k-fold cross-validation to evaluate the model’s performance. The results indicate that the model achieves acceptable error rates with reduced computational cost. Furthermore, the model was tested on a mobile device application, achieving a performance of 30 frames per second (FPS). The proposed system demonstrates potential for application in personalized therapy for patients and users.