Abstract

The Timed Up and Go (TUG) test is widely used for mobility assessment, but its clinical utility is limited by reliance on total completion time alone. Detailed analysis of constituent subtasks can provide richer insights into functional impairments. This study proposes a Bidirectional Long Short-Term Memory (BiLSTM) network with multilabel classification for robust activity segmentation of the Instrumented TUG (iTUG) test using a single lumbar-mounted IMU. Data from 105 participants (79 healthy across three age groups and 26 with gait-affecting pathologies) were collected. The model achieved a macro-average F1-score of 0.94 across all subtasks (sit-to-stand, walking phases, turns, and stand-to-sit). Clinical validation showed strong agreement with manual annotations (MAE: 0.28-0.40 seconds across subtasks). Significant between-group differences were found in all subtask durations, with the clinical cohort showing consistent prolongations. This work presents and clinically validates an automated iTUG segmentation method that supports detailed mobility assessment in clinical settings.