This study discusses Parkinson’s disease (PD) diagnosis from gait signals when data is limited and patients at new sites are different from those seen in training. The authors build a multi-channel 1D CNN that takes vertical ground-reaction force data from shoe-embedded sensors and pre-train it with a dynamic federated learning scheme on multiple public gait datasets. They then transferred the federated global model to a real-world population collected with a wearable insole system. Compared with direct training or plain federated pre-training, the combined approach significantly improves generalization to unseen subjects and sites. This achieved about 84% accuracy with balanced precision and recall in cross-subject testing on their XJTU dataset, and outperformed recent federated/transfer baselines that degrade on gait data. They also probed for effects like record-level splits or augmentation before splitting, which can indicate how such leakage can artificially boost scores, advocating strict cross-subject evaluation.