Changes in walking patterns are a well-known symptom of Parkinson's disease (PD), traditionally assessed through clinical observation. Advances in wearable technology and machine learning (ML) offer new opportunities for early detection, disease monitoring, and treatment evaluation. This study used wrist-worn accelerometry from the PPMI Verily study, collected from the first seven days of wear, to classify PD status. Raw accelerometer data were segmented into 10-second windows and filtered for active walking using activity recognition models. These walking windows were then classified into labels of healthy control, prodromal PD, or diagnosed PD using several ML models, to determine the best classifier. A median of 11 hours of active walking data per participant was analysed across 314 individuals: 34 healthy controls, 147 prodromal PD, and 133 diagnosed PD. Using 10-fold group cross-validation, the best-performing was a modified 18-layer ResNet-V2 with self-supervised pre-training, which achieved a mean (± standard deviation) macro F1 score of 0.59 ± 0.08 and AUROC of 0.86 ± 0.06. These findings highlight the potential of wrist-worn accelerometers powered by deep learning to identify PD-related gait changes. Future work may explore evaluating gait over longer monitoring periods or developing survival models for early risk prediction using an accelerometer-derived gait score.