Abstract
Purpose: This study aimed to develop feature extraction strategies for Center of Pressure (CoP) signals using adaptive genetic programming to characterize fall risk in older adults.
Methods: The individual performance of CoP indices reported in the state-of-the-art was optimized through adaptive genetic programming across mathematical domains, such as entropy, time-based (distance, area, hybrid measures) and frequency-based ones. The validity of the new CoP indices was tested using mean difference tests for groups with and without fall risk, measuring the correlation with existing measures, as well as through the performance of univariate and multiple logistic regressions, which were reported in terms of the macro-average F1-score, recall, accuracy, specificity, sensitivity, and Area Under the Curve (AUC).
Results: The newly generated genetic CoP indices outperformed state-of-the-art indices in fall risk identification. The genetic-frequency CoP index achieved the best performance in univariate logistic regression, with an AUC of 0.763 using five-fold cross-validation. Moreover, all genetic indices showed statistically significant differences between older adults with and without fall risk.
Conclusions: These results suggest that the proposed methodology provides some simple calculation formulas that facilitate its future adoption in clinical settings and increase fall risk classification performance by up to 27.0%.