Three-Class Parkinson’s disease screening from smartwatch motion signals: A comparison of physics-informed handcrafted features and deep embeddings with patient-level validation

Early and reliable screening for Parkinson’s disease (PD) remains challenging, particularly when distinguishing PD from other movement disorders with overlapping motor manifestations. This study compares two three-class wearable-based screening frameworks using the PADS dataset, comprising 469 participants and 5159 smartwatch measurement sessions. The first framework uses a CNN–demographic network to extract 128-dimensional embeddings from motion signals and demographic variables, followed by gradient-boosting classification. The second uses 150 physics-informed handcrafted features spanning time-domain, frequency-domain, tremor-specific, and multiscale wavelet descriptors, classified by a soft-voting XGBoost–LightGBM ensemble. Both frameworks were evaluated using five-fold patient-disjoint cross-validation with dataset-wide out-of-fold predictions. The handcrafted ensemble achieved 72.28 % ± 3.19 % patient-level accuracy, a macro F1-score of 0.662 ± 0.037, and a pooled macro ROC-AUC of 0.825. The final no-attention deep pipeline achieved 68.9 % ± 4.6 % accuracy, a macro F1-score of 0.618 ± 0.042, and a macro ROC-AUC of 0.799. Although the handcrafted framework produced numerically higher results, paired patient-level tests showed no statistically significant differences between the two pipelines. Ablation analysis indicated that frequency-domain features provided the largest unique contribution to the handcrafted model, while demographic fusion was the most influential component of the deep architecture and attention provided no consistent benefit. Out-of-fold SHAP and failure-case analyses showed that errors were concentrated at the PD–other movement disorder boundary, including both borderline and high-confidence misclassifications. These findings support smartwatch-based three-class screening and decision support, while highlighting the limitations of single-modality wrist-motion data for reliable differential diagnosis. External multimodal validation is required before the framework can support clinical deployment.