Enhancing fNIRS-Based Decoding: A Heterogeneous Ensemble Network for Six-Class Movement Classification

Abstract

Neurological upper-limb impairments severely affect functional independence, motivating the development of brain-computer interface (BCI) systems for decoding motor intention. Although functional near-infrared spectroscopy (fNIRS) is a promising modality, its low signal-to-noise ratio and large inter-subject variability make multi-class motor decoding difficult. In this study, a Heterogeneous Controlled Temporal Ensemble Network (HCTE-Net) is proposed to classify six different upper-limb motor tasks based on fNIRS signals. Twenty healthy participants were included for the data acquisition from the motor cortex region of the brain and standardized hemodynamic modeling was used for pre-processing of signals to include task-related dynamics. HCTE-Net combines temporal representations of LSTM, Bi-LSTM, and Peephole LSTM by probabilistic ensemble fusion. Among conventional machine learning models, XGBoost achieved the highest average accuracy of 69.68%, outperforming SVM (39.09%), LR (37.64%), DT (36.87%), RF (48.54%), LDA (39.39%), and NB (34.58%). The proposed HCTE-Net achieved a maximum accuracy of 95.46%, exceeding LSTM (84.35%), Bi-LSTM (87.17%), and Peephole LSTM (73.58%). The results of statistical analysis revealed significant differences among all methods (F = 659.324, p = 4.234 × 10⁻¹⁵²), and Tukey's HSD test showed significant improvement (p < 0.001). Experimental results demonstrate that HCTE-Net consistently outperforms single models across multiple evaluation metrics.