To investigate the clinical application of a deep-learning model based on six-degree-of-freedom (6DOF) lower-limb gait kinematics for bedside prediction of fall risk in hospitalized older adults.

This cross-sectional study enrolled 207 hospitalized older adults (65 to 93 years) from the orthopedic wards of Guangdong Provincial People’s Hospital. Inclusion criteria: age≥60 years, clear consciousness, and the ability to walk independently for at least 10 m. Patients unable to complete gait testing due to severe trauma or acute complications were excluded. Lower-limb 6DOF gait kinematic data were collected using a motion-capture system. According to the Morse fall scale (MFS), participants were categorized into low-, medium-, and high-risk groups. Gait time-series data were transformed into two-dimensional gait maps and analyzed using an attention-enhanced one-dimensional convolutional neural network (1D-CNN). Model performance was evaluated using accuracy, precision, recall, and F1 score, and further assessed with receiver operating characteristic (ROC) curves and confusion matrix analysis.

A total of 207 participants were included, comprising 59 high-risk, 72 medium-risk, and 76 low-risk individuals. On the independent test set (n=41), the model achieved an accuracy of 0.878, precision of 0.874, recall of 0.897, and an F1 score of 0.882. The recall for the high-risk group reached 1.000, indicating no missed high-risk fallers. ROC analysis demonstrated good discriminative ability across different risk levels.

The attention-enhanced 1D-CNN model based on 6DOF gait kinematics can effectively stratify fall risk in hospitalized older adults with high predictive performance and clinical interpretability. This approach provides an objective bedside tool for fall-risk screening in hospital settings.