DeepPIGD

1. Overview DeepPIGD provides action-separated depth-camera skeleton recordings and clinical annotations for five items of the Movement Disorder Society–Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), Part III relevant to postural instability and gait difficulty (PIGD). The dataset supports skeleton-based severity assessment and longitudinal change analysis across cross-sectional, inpatient and home cohorts. It includes anonymized participant characteristics, recording metadata, item-level clinical scores and baseline-relative change labels. 2. Participants and Dataset Composition DeepPIGD includes 50 distinct participants, comprising 25 male and 25 female participants, aged 43–91 years (mean: 68.08 years). Recorded disease duration ranges from 0 to 20 years. Anonymized participant information includes sex, age, disease duration, Hoehn–Yahr stage and reported collection setting. The cohorts overlap; their participant counts should therefore not be summed. Cohort Participants Analysis units Recording takes Action-level files Follow-up comparisons Cross-sectional 48 300 records 300 888 — Inpatient longitudinal 11 58 assessments 115 342 47 Home longitudinal 3 18 visits 39 117 15 Total 50 unique 376 454 1,347 62 For the cross-sectional cohort, each record corresponds to a recording take. Its clinical labels are inherited from the participant-level assessment and do not represent an independent clinical rating for each take. Longitudinal analysis units correspond to labeled assessments or visits, each of which may include multiple recording takes. 3. Data Modality and Format The release contains Azure Kinect depth-derived skeletons in action-separated NumPy object files. Each detected body contains 32 joints, with: Three-dimensional positions in millimeters, expressed in the camera coordinate system. Joint orientations. Joint confidence levels. The existing action-separated skeleton files are preserved without further resampling, normalization or frame removal. Original frame sequences and detected-body lists are retained. The nominal acquisition rate is 30 Hz; per-frame timestamps are not included. Reading the original object files requires compatible NumPy and PyKinectAzure dependencies. The supplied example loader also supports exporting numeric arrays for subsequent analysis. 4. Motor Tasks and Longitudinal Structure Recordings cover three motor tasks: sit-to-stand, standing, and walking-turning. The release contains 449, 444 and 454 action-level files for these tasks, respectively. Each recording take contributes up to three action files; missing actions are not imputed. The cross-sectional cohort includes 471 single-task (ST) and 417 dual-task (DT) files. The inpatient and home cohorts contain 342 and 117 ST files, respectively. Eleven additional home skeleton files without matching clinical ground truth are excluded. Within each longitudinal participant and cohort, day0, day1, day2. Each labeled record receives a separate index. Absolute visit dates and actual time intervals are not distributed. 5. Clinical Annotations and Participant Metadata Clinical ground truth covers the following MDS-UPDRS Part III items: Item Clinical measure 3.9 Arising from chair 3.10 Gait 3.11 Freezing of gait 3.12 Postural stability 3.13 Posture Original integer scores of 0–4 are preserved, together with their five-item sum (0–20). Separately named model-label columns apply min(score, 3), merging scores 3 and 4 and yielding a derived total of 0–15. These columns support reproduction of the four-class modeling convention; they are not additional clinical assessments. No model predictions are used as ground truth. The metadata also include participant demographics, cohort membership and recorded medication on/off status for all 58 inpatient assessments. Medication status is left blank where unavailable. Participant-level Hoehn–Yahr stage is provided as a clinical characteristic, not as a visit-specific longitudinal measurement. All longitudinal comparisons use the participant’s first assessment in the same cohort as the baseline. Total-score changes are classified as: Improved: change < −1. Stable: −1 ≤ change ≤ 1. Worsened: change > 1. Comparisons based on original clinical scores and derived model-label scores are supplied separately. An anonymized annotation note documents one source-table total-score discrepancy; the released total follows the sum of the five original item scores. 6. Research Applications DeepPIGD supports research on: Skeleton-based ordinal estimation of PIGD-related clinical scores. Baseline-relative longitudinal change detection. Comparisons between single-task and dual-task recordings. Assessment across inpatient and home recording settings. Consistent anonymous participant identifiers enable participant-level grouping across cohorts when constructing evaluation splits. Because participants overlap between cohorts, cohort membership alone does not ensure participant-independent evaluation. 7. Files, Anonymization and Access The dataset is distributed in four ZIP archives: cross-sectional skeletons, inpatient skeletons, home skeletons and metadata. Extract all four archives into the same directory. The metadata archive includes participant characteristics, session and recording-take tables, clinical labels, file manifests, joined annotations, longitudinal comparisons, joint definitions, annotation notes, a data dictionary, dependencies, an example loader and a verification utility. SHA-256 checksums are provided for both the downloadable archives and the extracted files. Public identifiers are anonymized and consistent across cohorts. Original names, source participant identifiers, original directory paths and absolute visit dates are excluded from public metadata and filenames. The private identity mapping is not distributed. Skeleton files retain all detected bodies. The example loader selects the first detected body in each frame; this convention does not guarantee persistent person tracking when multiple bodies are present. Access: Record metadata are public. Dataset files are restricted and available to users granted access by the record owner. License: Creative Commons Attribution 4.0 International (CC BY 4.0). Related publication: The associated manuscript is currently unpublished, and no related publication DOI is assigned to this record.