YannisTevissen/sma-upper-limb-kinect

SMA Upper-Limb Kinect Dataset and Reach-Intent Benchmark

This repository contains the four supplementary files from a longitudinal Kinect study and a derived reach-intent benchmark.

Publication files

The repository contains the four supplements listed by PLOS:

• pone.0170472.s001.pdf: complete analysis report;
• pone.0170472.s002.zip: prototype game, R analysis, and Python extraction source code;
• pone.0170472.s003.zip: raw data, logs, extracted features, and demographics;
• pone.0170472.s004.mp4: animated replay of an outlier recording.

The accompanying MANIFEST.csv records the DOI, byte size, SHA-256 checksum, licence, and description of each file.

The question in plain language

A person sees one of ten targets and reaches toward it. Given only the beginning of their 3D upper-body movement, can a model guess the intended target?

The target is not whether a participant has SMA. SMA/control labels are included only to check whether a system works comparably across groups.

Data

The study contains approximately 16 Hz 3D coordinates for nine joints, task logs, and repeated sessions from 18 ambulant adults with SMA type III and 19 controls. The derived tables under data/ join the logs to the skeleton stream and exclude the clinical table and fine-grained demographics.

• data/reach_trials.csv.gz: one row per target presentation.
• data/reach_frames.csv.gz: skeleton frames from target appearance to reach or timeout.
• data/sessions.csv: coarse session metadata, visit order, and paper-QC flag.
• data/participant_folds.csv: fixed participant-disjoint five-fold assignment.
• data/build_summary.json: counts and source checksum.

The source contains 197 raw filenames and 197 log filenames, but one entry in each set has no matching counterpart. The build does not guess a match. It also reproduces the public analysis code's cohort-quality rules: participant 1018 is removed because the source analysis marks them “not SMA,” participant 1036 is absent from the clinical study table, the 1027 recording from 10 July 2014 is marked “no real game,” and non-study session dates are filtered. The paper reports 187 valid feature records (66 SMA and 121 control); two valid control records lack one side of the exact raw/log pair, leaving 185 paired sessions (66 SMA and 119 control) in the downloadable tables. The two later PCA outlier sessions are retained with qc_outlier=true; excluding them leaves 183 paper-QC sessions for the primary benchmark. All exclusions and flags are recorded in build_summary.json.

Coordinates preserve the source column names and units. progress runs from 0 near target appearance to 1 at the recorded reach/timeout endpoint. target_label is the task label from 0 to 9. Objects 10–19 are the second presentation of positions 0–9.

The game starts object 0 without writing an appeared event. For its first presentation only, the builder uses the first available raw skeleton timestamp as the observable start and sets start_inferred=true. Other presentations use explicit event timestamps. Report sensitivity with these inferred-start trials removed when evaluating very early intent.

Benchmark protocol

Use completed reaches with qc_outlier=false for the primary benchmark; timed-out attempts and known QC outliers are optional robustness sets.

1. Generic: train on four participant folds and test on the held-out fold.
2. Personalized 1-shot: use the first completed example of each target from the test participant as calibration, then test on their later examples.
3. Personalized 5-shot: same protocol with the first five examples per target.
4. Cross-visit: calibrate on visit 0 and test only on later visits.

Evaluate intent at 25%, 50%, and 100% of the movement. Report overall accuracy, macro target recall, and accuracy separately for the SMA and control slices. Never split individual trials from one person across generic train and test sets.

Run the dependency-free nearest-centroid reference:

python scripts/evaluate_baseline.py --data-dir data

This baseline is a sanity check, not a state-of-the-art claim. The personalized comparison evaluates generic, personal-only, and prior-weighted adapted centroids on the same post-calibration trials, so adapted_accuracy_gain is a fair within-test-set comparison.

data/baseline_results.json contains the reference output for the paper-QC primary set, the stricter explicit-start sensitivity slice, and all exact raw/log pairs.

Source citation and attribution

• Study: Chen et al., Feasibility of Using Microsoft Kinect to Assess Upper Limb Movement in Type III Spinal Muscular Atrophy Patients, PLOS ONE (2017).
• Article DOI: 10.1371/journal.pone.0170472
• Data supplement: PLOS S3 dataset
• Source code supplement: PLOS S2 code

Chen X, Siebourg-Polster J, Wolf D, Czech C, Bonati U, Fischer D, Khwaja O, Strahm M. (2017). Feasibility of Using Microsoft Kinect to Assess Upper Limb Movement in Type III Spinal Muscular Atrophy Patients. PLOS ONE, 12(1), e0170472. https://doi.org/10.1371/journal.pone.0170472

The paper and supplements are distributed under CC BY 4.0. Redistribution requires credit to the original authors and source. When using a supplement, cite the paper and the corresponding supplement DOI; when using the derived files, also identify this Hugging Face repository version.

Limitations and responsible use

This is a small cohort of ambulant, relatively high-functioning adults with SMA type III. It does not represent the full SMA spectrum, children, non-ambulant people, or deployment conditions. Kinect tracking and the fixed game interface introduce measurement bias. Do not use these files for diagnosis, prognosis, treatment decisions, identity inference, or claims of clinical validity.