This paper proposes a computational modeling oriented framework for intelligent knee function assessment and movement risk prediction in elderly patients. The framework revolves around the joint modeling of gait video, inertial signals, plantar pressure sequences, and knee flexion and extension measurements, and no longer relies on single source observations. This study constructs a data set containing 1240 evaluation samples from 186 elderly patients, and integrates multi-modal feature fusion, temporal representation of functional status and probabilistic risk inference mechanism in a unified framework to characterize the changes in joint stability, mobility and motor coordination. The experimental results show that the mean absolute error of functional score estimation is 4.9%, the accuracy of risk prediction is 91.3%, and the precision, recall and F1-score are 92.1%, 89.4% and 90.7%, respectively. Compared with CNN-LSTM, random forest and unimodal baseline models, the proposed framework shows strong stability and robustness under complex continuous action clips of elderly patients, which can provide feasible technical support for digital rehabilitation assessment and risk early warning in intelligent medical scenarios.
