Objective To predict the torque on the affected side of the hip, knee, and ankle joints in stroke patients during walking using principal component analysis (PCA) and backpropagation (BP) neural networks. Methods Kinematic and dynamic data from 30 stroke patients were synchronously collected using an 8-lens Qualisys infrared point high-speed motion capture system and Kistler three-dimensional (3D) force measurement platform. The torques of the hip, knee, and ankle joints in the stroke patients were calculated using OpenSim, and the initial variables with a cumulative contribution rate of 99% were screened using PCA. The normalized root mean square error (NRMSE), root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), and R2 were used as evaluation indicators for the PCA-BP model. The consistency between the calculated joint torque and predicted torque was evaluated using Kendall's W coefficient. Results PCA data showed that the trunk, pelvis, and affected sides of the hip, knee, and ankle joints had a significant impact on the torque of the affected sides of the hip, knee, and ankle joints on the x, y, and z axes (sagittal, coronal, and vertical axes, respectively). The NRMSE between predicted and measured values was 5.14%―8.86%, RMSE was 0.184―0.371, MAPE was 3.5%―4.0%, MAE was 0.143―0.248, and R2 was 0.998―0.999. Conclusions The established PCA-BP model can accurately predict the torque of the hip, knee, and ankle joints in stroke patients during walking, with a significantly shortened measurement time. This model can replace traditional joint torque calculation in the gait analysis of stroke patients, provides a new approach to obtaining biomechanical data of stroke patients, and is an effective method for the clinical treatment of stroke patients.