INTRODUCTION: The Timed Up and Go dual-task test (TUG-DT) is a brief probe of cognitive-motor interference, but its screening value is limited by manual scoring and underuse of concurrent speech performance.

METHODS: We developed ADualNet, an interpretable hybrid-AI framework that extracts pose-based gait features and spoken serial-subtraction measures from synchronized TUG-DT video and audio. In 164 Shenzhen Multimodal Aging Research (STAR) participants, models were evaluated using leakage-controlled nested feature selection and repeated cross-validation, with domain ablation, calibration, model interpretation and continuous MoCA regression.

RESULTS: The best-performing AdaBoost model achieved an AUROC of 0.893 (95% CI, 0.877–0.907), accuracy of 0.808, precision of 0.839, recall of 0.747 and F1 score of 0.786. The strongest model combined gait, speech-task and demographic variables. Speech-task measures contributed more than gait measures within the behavioral domains. The extracted feature space also predicted continuous MoCA scores (R² = 0.489).

DISCUSSION: ADualNet converts brief TUG-DT recordings into interpretable risk profiles for first-line cognitive impairment screening. By automating dual-task feature extraction and risk stratification, it may support outpatient, community and remote triage, helping identify individuals who warrant further neuropsychological or biomarker-based evaluation. External and longitudinal validation is needed before clinical implementation.