Multimodal Machine Learning Approach for Timed-Up-and-Go Test Score Prediction from Wearable Gait Data

Falls are a major cause of injury and disability in older adults and people with Parkinson’s disease (PD). The Timed-Up-and-Go (TUG) test assesses mobility and falls risk but requires supervised clinical testing, limiting its regular use. We investigated whether TUG time can be estimated from two-minute walking data captured using in-shoe vertical ground reaction force (GRF) sensors, which are easy to use during daily life and provide detailed gait information. While prior studies have estimated TUG from wearable sensors, how different input modalities such as gait representations and demographics contribute individually and in combination has received little attention.

Using the PhysioNet GRF dataset, we analyzed 100 participants (62 PD, 38 controls) with demographics (age, sex, height, and weight) and 16-channel GRF recordings. A Perceiver-based multimodal framework was trained to estimate TUG time from raw GRF timeseries, GRF-derived summary features, and demographics in unimodal and multimodal configurations and benchmarked against conventional feature-based regressors trained on extracted features. Models were evaluated using ten-fold subject-wise cross-validation with mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (CC).

Combining raw timeseries and features (MAE 0.80 ± 0.34 s, RMSE 1.27 ± 0.85 s, CC 0.93 ± 0.07) achieved best performance, outperforming unimodal models, conventional regressors, and prior GRF-based TUG results. PD participants showed higher prediction errors (MAE 0.97 s) than controls (MAE 0.41 s), with accuracy decreasing in older age groups. Adding demographics provided no consistent benefit.

These findings suggest that steady-walking vertical GRF can approximate TUG time and could support continuous mobility assessment in PD.