MobileNetV2 for Multiclass Neurodegenerative Gait Analysis Using Bilateral Ground Reaction Force Images: From Transfer Learning to Raspberry Pi 4 Deployment

Neurodegenerative diseases (NDDs), including Parkinson’s disease (PARK), Huntington’s disease (HUNT), and amyotrophic lateral sclerosis (ALS), are associated with characteristic gait alterations. This study proposes an image-based deep-learning approach for NDD gait classification and its implementation on an edge-computing platform. The main contributions include: (1) mapping synchronized left- and right-foot ground reaction force (GRF) signals from individual gait cycles into a common 2D bilateral force space to generate three complementary representations—Trajectory Map (bilateral force trajectories), Derivative Map (dynamic transitions), and Occupancy Heatmap (loading-state distributions); (2) adapting ImageNet-pretrained MobileNetV2 to four-class gait classification (CONTROL, ALS, HUNT, and PARK) through classification-layer retraining and feature-extractor fine-tuning; (3) comparing compact (α = 0.35; ~0.4 M parameters) and higher-capacity (α = 1.4; ~4.4 M parameters) MobileNetV2 configurations; and (4) evaluating the deployment of the resulting models on a Raspberry Pi 4 as a resource-constrained edge-computing platform. The models were evaluated on a test subset of 2259 gait cycles from the Gait in Neurodegenerative Disease Database, comprising CONTROL (612), ALS (367), HUNT (727), and PARK (553). Fine-tuning increased overall four-class accuracy from 89.64–92.61% to 95.09–96.41% across the GRF representations and model configurations. Paired comparisons showed improvements of 3.81–6.29 percentage points across all six matched configurations. Class-specific one-vs-rest accuracy was consistently high after fine-tuning: CONTROL (97.74–98.63%), ALS (98.41–98.94%), HUNT (96.02–97.21%), and PARK (97.30–98.23%). The highest four-class accuracy (96.41%; 95% CI: 95.45–97.29%; Occupancy Heatmap, α = 1.4) was preserved for the unquantized TFLite model evaluated on the Raspberry Pi 4, while dynamic-range quantization reduced accuracy by 0.22 percentage points to 96.19%. Compact fine-tuned MobileNetV2 models required approximately 42–46 ms of inference latency and 89–97 mJ of estimated incremental energy per inference, compared with 324–374 ms and 680–785 mJ, respectively, for the higher-capacity models. Dynamic-range quantization reduced model storage by approximately 62% (α = 0.35) and 72% (α = 1.4). Overall, bilateral GRF representations combined with MobileNetV2 transfer learning enable accurate multiclass NDD gait classification, while compact models provide a favorable performance–efficiency trade-off for resource-constrained edge implementation. However, subject-independent validation remains necessary because the reported accuracy was obtained using gait-cycle-level data partitioning.