Objectives: Accurate severity classification is important for sensor-based assessment of Parkinson’s disease, but overlap between adjacent stages and class imbalance can reduce model robustness. This study aimed to develop and evaluate a dual-task gait fusion framework that integrates signals collected during self-selected walking and walking with turning. Methods: The primary cohort comprised 87 participants with Parkinson’s disease across mild, mild-to-moderate, and moderate stages. Task-specific temporal representations were learned from the two gait conditions, concatenated and optimized using delayed class re-weighting. Performance was evaluated using subject-level stratified five-fold cross-validation with five random seeds. Generalizability was further assessed using an independent PhysioNet cohort of 93 participants with Parkinson’s disease. Results: The proposed method achieved an accuracy of 90.23 ± 1.27, balanced accuracy of 89.61 ± 1.13, macro F1-score of 89.10 ± 1.35, and probability-based macro AUC of 94.43 ± 1.42 on WearGait-PD. Confusion matrix and ablation analyses indicated balanced class-level performance and complementary contributions from dual-task fusion and delayed re-weighting. On the external PhysioNet cohort, accuracy, balanced accuracy, macro F1-score, and macro AUC were 84.34 ± 1.82, 82.13 ± 1.92, 81.68 ± 1.98, and 88.93 ± 2.14, respectively. Conclusions: Integrating turning-related gait information with self-selected walking signals improved wearable sensor-based severity classification and showed cross-dataset robustness, although validation in larger cohorts with complete severity stage coverage remains necessary.

Wearable inertial sensors offer a portable means of assessing gait abnormalities associated with neurological and orthopedic disorders. However, intersubject variability, overlapping gait patterns, and class imbalance complicate multiclass recognition. This study presents the Temporal Structural Dual Graph Network (TSDGNet) for pathological gait classification using inertial measurement unit (IMU) signals. Four IMUs positioned on the head, lower back, and both feet provide acceleration and angular velocity signals during walking and turning. The framework combines multiscale temporal feature extraction with graph-based modeling of sequence-level relationships among sensor representations. Two fixed prior graphs encode anatomical and symbiotic relationships, while trainable graph attention and context-guided gated residual refinement adapt sensor-node representations. Class-balanced optimization is incorporated to reduce training bias toward majority classes. Subject-independent evaluation was performed on a public clinical dataset comprising healthy controls and seven neurological or orthopedic disease groups. Comparisons with representative baseline and advanced models showed competitive classification performance. Ablation analyses examined the contributions of temporal modeling, structural relationship learning, and class-balanced optimization. These findings support the potential of combining wearable inertial sensing with temporal and graph-based learning for objective classification of pathological gait.