AUT Journal of Electrical Engineering

AUT Journal of Electrical Engineering

Improved Metric for Classification of Nearby Reaching Targets: A Distance-Weighted Accuracy Approach

Document Type : Research Article

Authors
1 Faculty of Biomedical Engineering, Amirkabir University of Technology, Tehran, Iran
2 Faculty of Biomedical Engineering, Semnan University, Semnan, Iran
10.22060/eej.2026.25416.5930
Abstract
Accurate classification of reaching targets is essential for upper-limb prosthesis control, rehabilitation robotics, and human–robot interaction. Conventional accuracy assumes equal misclassification costs and therefore does not distinguish between errors involving spatially adjacent and distant targets. We propose a spatially informed distance-weighted accuracy metric that assigns misclassification costs according to the normalized Euclidean distance between intended and predicted targets, with greater penalties for more distant errors. We first demonstrate the utility of the proposed metric using synthetic confusion matrices with identical conventional accuracy (33.3%) but distinct spatial error patterns (near, far, and random). While conventional accuracy remained 33.3% for all three cases, distance-weighted accuracy yielded scores of 77%, 33.3%, and 49%, respectively with the weighting parameter (ω) set to 1. The proposed metric was also evaluated on a real-world EEG-based reaching-target classification task using Quadratic SVM (QSVM) and Gaussian SVM (GSVM) classifiers. Performance was assessed using five repetitions of two-fold cross-validation (5×2 CV), and classifier differences were evaluated using the combined 5×2 CV F-test. Conventional accuracy was 48.90 ± 6.33% for QSVM and 31.30 ± 8.24% for GSVM, with a statistically significant difference (p = 0.00476). Using the proposed distance-weighted accuracy with , QSVM achieved 87.12 ± 1.75%, compared with 75.93 ± 9.44% for GSVM. however, this difference was not statistically significant (p = 0.197). These findings demonstrate that the proposed metric provides a complementary, distance-aware assessment of classification performance by accounting for the physical distances between predicted and true classes beyond conventional exact-match accuracy.
Keywords
Subjects


Articles in Press, Accepted Manuscript
Available Online from 21 November 2026