To address the demand for mouse-equivalent access among people with upper-limb impairments, numerous intelligent interaction frameworks have been proposed. However, existing systems still exhibit significant limitations in terms of perception dimensionality, mapping accuracy, and personalized adaptation. Therefore, starting from the design of the fundamental sensing unit, a novel MXene/MWCNT/MXene sandwich-structured sensitive layer is designed and fabricated using a KOH-ion-induced gelation reaction combined with a layer-by-layer vacuum filtration process. Based on this sandwich-structured sensitive layer, a flexible piezoresistive sensor is assembled by integrating interdigital electrodes and a PDMS encapsulation layer. Benefiting from the robust conductive pathways, the developed sensor exhibits 10.97 kPa-1 sensitivity and 100 ms fast response time, as well as below 5% drift over 800 loading cycles, demonstrating excellent mechanical and electrical stability. On this basis, the proposed sensor is integrated with an inertial measurement unit (IMU) and a rotary potentiometer to construct an intelligent interactive control system that fuses multimodal information, including pressing, rotation, and displacement. The system employs a sliding-window-based multi-scale feature extraction method to construct a fused feature vector that simultaneously captures local dynamic characteristics and global steady-state features. This fused representation is then processed by a graph neural network (GNN), achieving an average recognition accuracy of 97.2% across 12 typical mouse interaction actions. Meanwhile, a user-specific thresholding method that tracks individual behavior vectors boosts overall accuracy by 5-4% while reducing false-trigger rates by 30%. The proposed approach offers a highly robust interaction solution for users with hand-function impairments and lays the groundwork for personalized adaptation in intelligent human-computer interaction systems.