Automated assessment of balance function based on a lightweight end-to-end multimodal deep learning framework: predicting Mini-BESTest scores from center of pressure signals in a single stance condition

Background Balance dysfunction is a common health issue among older adults and patients with neurological disorders. The Mini-Balance Evaluation Systems Test (Mini-BESTest) is a widely used clinical tool for balance assessment. However, its full implementation requires a trained clinician to administer 14 tasks, and the scoring is subject to inter-rater variability. Objective This study aims to develop an end-to-end, lightweight multimodal deep learning framework (with approximately 55 K parameters) to automatically predict the total Mini-BESTest score and its four subscale scores, utilizing only center of pressure (COP) signals from a single stance condition alongside clinical metadata. Methods COP signals and clinical metadata from 163 subjects in a public dataset (Santos et al., PeerJ, 2016, 4) were utilized. We proposed a lightweight multimodal deep learning model comprising three dilated convolution branches with varying dilation rates, a channel attention mechanism, and a metadata fusion module. Three independent comparative experiments were conducted to investigate the effects of stance condition, trial number, and time window length on predictive performance. Based on these results, a suitable data configuration was selected for subsequent experiments. An adaptive task weighting scheme based on homoscedastic uncertainty was employed to balance the five regression targets with varying score ranges. Five-fold cross-validation was used to evaluate the models, and out-of-fold (OOF) predictions were constructed to provide a full-sample assessment of predictive performance. Results Under the selected data configuration (foam surface with eyes open, second trial, first 30 s of data), the five-fold cross-validation mean absolute error (MAE) for the predicted Mini-BESTest total score was 1.971 (out of 28). The full-sample OOF prediction MAE was 1.979, with a Pearson correlation coefficient of r = 0.758 and R 2 = 0.574. In ablation experiments, removing the metadata branch led to the largest numerical increase in total-score MAE (to 2.338), while removal of the channel attention module resulted in a smaller increase (to 2.034). A metadata-only model achieved a total-score MAE of 2.105. Paired Wilcoxon signed-rank tests did not show statistically significant differences between the full model and the ablated variants (all p > 0.05). Conclusion This study demonstrates that an end-to-end deep learning framework can predict the complete Mini-BESTest score from 30 s of quiet standing COP signals under foam surface eyes-open conditions, together with basic clinical metadata. While maintaining predictive accuracy, this lightweight framework (approx. 55 K parameters) may represent a potential direction for future deployment in portable or remote settings, subject to further engineering and clinical validation.