Explicit Boundary Formation for Safe and Predictable Head-Movement Classification

Safety‑critical applications that use classification algorithms must not only classify accurately but also behave predictably when confronted with inputs outside the distribution of the original training data. Artificial Neural Networks (ANNs), while powerful, form decision boundaries implicitly and often extrapolate or generalise unpredictably when provided with such inputs. This unpredictability poses significant risks in assistive‑technology interfaces such as head‑movement control for powered wheelchairs, where misclassification can lead to unsafe or unintended motion. This paper re‑examines the generalisation problem from a boundary‑centric perspective and introduces the Optimised Region and Boundary Classifier (RBC), a geometric classifier that constructs explicit, interpretable decision regions using hyper‑rectangles and null‑region exclusions. Using real head‑movement data from able‑bodied and spinal‑injured participants, we compare the RBC with ANNs trained on conventional and augmented datasets, as well as with an optimised Hyper‑Rectangle Basis Function (HRBF). Performance is evaluated using the 0.632+ bootstrap estimator and Receiver Operating Characteristic (ROC) analysis. Results show that representative training data substantially improves ANN performance, while the RBC achieves the highest overall performance across all evaluated metrics. These findings demonstrate that explicit, optimised boundaries can offer a principled and reliable alternative to implicit ANN decision surfaces in safety‑critical applications.