tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

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

Authors :

Zahra Dayani¹ Ali Maleki² Ali Fallah³

1- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران) 2- دانشگاه سمنان 3- دانشگاه صنعتی امیرکبیر(پلی تکنیک تهران)

Keywords :

reaching target classification،upper-limb prosthesis control،spatially weighted accuracy،performance evaluation metrics،misclassification cost،motor intention decoding

Abstract :

Accurate classification of reaching targets is critical for upper-limb prosthesis control, rehabilitation robotics, and human-robot interaction. Traditional classification metrics assume uniform misclassification costs, ignoring the spatial relationships between targets. This overlooks significant performance degradation: misclassifications in safety-critical zones (e.g., near obstacles or humans) or those impairing functional outcomes (e.g., failing to grasp a cup) can be far more detrimental than spatially adjacent misclassifications—despite equivalent cost in standard metrics—leading to elevated user workload or complete task failure. To address this, we propose a spatially informed weighted accuracy metric. Misclassification costs are assigned based on the normalized Euclidean distance between the intended target and the misclassified position, penalizing distant errors more heavily than proximal ones. We demonstrate the utility of this metric first using synthetic confusion matrices achieving identical standard accuracy but exhibiting distinct spatial error patterns (far, near and random misclassification error patterns). We then apply it to a real-world reaching target prediction task, comparing two classifiers (Quadratic Kernel SVM vs. Gaussian Kernel SVM) with equal standard accuracy (63%). The proposed metric effectively discriminates classifier performance by imposing higher penalties on distant misclassifications (86.3% for Quadratic Kernel SVM vs. 85.5% Gaussian Kernel SVM), revealing significant differences masked by standard accuracy. Crucially, the metric explicitly normalizes against the worst-case misclassification cost inherent to the target layout, providing a spatially aware assessment of classification performance essential for real-world deployment.