Gait-Based Machine Learning Model for Detecting Prevalent Mild Cognitive Impairment in Older Chinese Adults: A Community-Based Cross-Sectional Study.

INTRODUCTION

Detection of prevalent mild cognitive impairment (MCI) is difficult. We tested whether adding gait features to clinical variables improved prevalent MCI classification among older adults.

METHODS

This pooled cross-sectional analysis included 266 older adults. A Kinect sensor was used to capture skeletal trajectories, from which spatiotemporal, variability, asymmetry, center-of-gravity stability, and kinematic range-of-motion features were computed. MCI was assessed by Petersen's criteria. Machine learning algorithms, trained using three predictor sets (clinical-only, gait-only, and a fused set combining both), were employed to predict prevalent MCI, with performance evaluated primarily using the area under the ROC curve (AUROC), supplemented by the area under the PR curve (AUPRC), etc. SHapley Additive exPlanations (SHAP) was used to identify key predictors.

RESULTS

The study included 76 MCI and 190 cognitively normal participants (mean age: 74.9 years; 73.2% women). The optimal model for clinical-only set was neural network, achieving an AUROC of 0.687 (SD: 0.081) and an AUPRC of 0.526 (SD: 0.111). For gait-only set, logistic regression performed best, with corresponding values of 0.813 (SD: 0.060) and 0.636 (SD: 0.122). Feature fusion yielded the best performance: LightGBM model attained an AUROC of 0.819 (SD: 0.070) and an AUPRC of 0.656 (SD: 0.118), with an accuracy of 0.776 (SD: 0.056), sensitivity of 0.596 (SD: 0.128), and specificity of 0.846 (SD: 0.057). SHAP analysis of the LightGBM identified step length asymmetry, step length variability, and velocity as key predictors.

CONCLUSION

Markerless depth camera gait assessment showed potential in identifying prevalent MCI, which needs to be validated in external cohorts.