A three-year cross-sectional study was conducted to develop a comprehensive dataset for machine learning (ML) models in Parkinson's disease (PD) research. The study involved 504 participants, including PD patients, differential diagnoses (DD), and healthy controls (HC), with over 5000 clinical assessment steps recorded using a multi-modal smartphone app and smartwatches. The dataset includes detailed annotations on demographics, medical history, symptoms, and movement steps, providing a rich source of data for ML analysis. An integrative ML approach combining classical signal processing and deep learning techniques was implemented and cross-validated. The models achieved an average balanced accuracy of 91.16% in classifying PD vs. HC and 72.42% in classifying PD vs. DD. The results suggest that the dataset has significant potential for improving diagnostic accuracy and monitoring treatment efficacy in PD. The study also highlights the importance of using multimodal data, including both movement data and self-reported non-motor symptoms, to enhance classification performance. The PADS dataset, which includes smartwatch and smartphone data, is publicly available for further research. The study emphasizes the need for large, balanced datasets to ensure generalizability and reduce overfitting. The findings demonstrate that the integration of different data modalities can improve diagnostic accuracy, particularly in distinguishing PD from similar movement disorders. The study also addresses potential biases, such as gender imbalance, and suggests that further research is needed to validate the results in diverse populations. Overall, the study contributes to the development of ML models for PD diagnosis and monitoring, with the potential to improve clinical outcomes through early detection and personalized treatment.A three-year cross-sectional study was conducted to develop a comprehensive dataset for machine learning (ML) models in Parkinson's disease (PD) research. The study involved 504 participants, including PD patients, differential diagnoses (DD), and healthy controls (HC), with over 5000 clinical assessment steps recorded using a multi-modal smartphone app and smartwatches. The dataset includes detailed annotations on demographics, medical history, symptoms, and movement steps, providing a rich source of data for ML analysis. An integrative ML approach combining classical signal processing and deep learning techniques was implemented and cross-validated. The models achieved an average balanced accuracy of 91.16% in classifying PD vs. HC and 72.42% in classifying PD vs. DD. The results suggest that the dataset has significant potential for improving diagnostic accuracy and monitoring treatment efficacy in PD. The study also highlights the importance of using multimodal data, including both movement data and self-reported non-motor symptoms, to enhance classification performance. The PADS dataset, which includes smartwatch and smartphone data, is publicly available for further research. The study emphasizes the need for large, balanced datasets to ensure generalizability and reduce overfitting. The findings demonstrate that the integration of different data modalities can improve diagnostic accuracy, particularly in distinguishing PD from similar movement disorders. The study also addresses potential biases, such as gender imbalance, and suggests that further research is needed to validate the results in diverse populations. Overall, the study contributes to the development of ML models for PD diagnosis and monitoring, with the potential to improve clinical outcomes through early detection and personalized treatment.