A Smart Insole-Based Explainable AI Framework for Non-Invasive Sciatica Diagnosis Using Optimally Selected Plantar Force Sensors

Abstract Sciatica is a neuropathic disorder that can cause significant alterations in lower-limb biomechanics, including gait asymmetry and abnormal plantar force distribution. This study proposes a wearable, explainable artificial intelligence (XAI)-based framework for automated sciatica diagnosis using optimally selected force-sensitive resistor (FSR) sensors integrated into a smart insole. A dataset of 37 subjects, comprising 20 patients with Sciatica and 17 healthy subjects, was acquired using 16 FSR sensors, with eight sensors positioned on each foot, along with concurrent electromyography (EMG) and inertial measurement unit (IMU) data. From the bilateral plantar force signals, 601 biomechanical features were extracted, incorporating statistical, temporal, magnitude, and inter-limb asymmetry characteristics. A K-fold combinatorial sensor-selection strategy identified the optimal sensor configuration, with K = 5 providing the most stable performance. The optimal configurations, L2, L5, L8 for the left foot and R2, R5, R8 for the right foot, reduced the number of sensors from 16 to 6, representing a 62.5% reduction in sensing hardware. The selected left-foot sensors achieved 96% accuracy, 96% precision, 93% recall, 94% F1-score, and 0.93 AUC, while the right-foot sensors achieved 87% accuracy, 83% precision, 76% recall, 79% F1-score, and 0.91 AUC. Using the six optimally selected sensors, the proposed one-dimensional convolutional neural network (1D-CNN) achieved 90.3% classification accuracy, 89.9% sensitivity, 90.3% specificity, and a ROC-AUC of 0.911. Furthermore, XAI-based plantar heatmaps confirmed the contribution of the selected biomechanical regions to the classification decision. Cross-validation against the concurrent EMG/IMU modality achieved 85.2% accuracy, with approximately 81.5% diagnostic agreement between the force-sensor and EMG/IMU models. The quantitative results demonstrate that a substantially reduced FSR sensor configuration can preserve diagnostically relevant biomechanical information while providing accurate, interpretable, and portable sciatica screening. The proposed framework offers potential for low-cost wearable gait assessment and automated clinical decision support.