Analysis code and derived data for: Demographic Shortcuts Account for Most of the Apparent Accuracy of Four-Class Gait-Based Differential Diagnosis in Neurodegenerative Disease: An Interpretability-Driven Audit of Machine Learning on a Public Gait Benchmark

Analysis code and derived data for the study "Demographic Shortcuts Account for Most of the Apparent Accuracy of Four-Class Gait-Based Differential Diagnosis in Neurodegenerative Disease: An Interpretability-Driven Audit of Machine Learning on a Public Gait Benchmark". This deposit contains everything generated by the study. It does not redistribute the source recordings, which remain permanently archived at PhysioNet. CONTENTS code/ (14 files) — every analysis script, numbered in execution order: download with SHA-256 verification, quality control and feature extraction, baseline and sequence models, SHAP interpretability, ablation and shortcut audit, the external cohort, and the figure builders. derived_data/ (19 files) — subject- and window-level feature tables, the stride-interval series extracted from the external cohort's force-plate recordings, out-of-fold predictions for every model and feature set, permutation null distributions, SHAP value arrays, and the cohort tables with the metadata used as confounders. result_tables/ (94 files) — every numeric result in the manuscript and its supplement, one CSV per analysis, including the quality-control logs and the corrected-versus-originally-computed comparisons. SOURCE DATA (not included) Gait in Neurodegenerative Disease Database (gaitndd v1.0.0), https://physionet.org/content/gaitndd/1.0.0/Gait in Parkinson's Disease Database (gaitpdb v1.0.0), https://physionet.org/content/gaitpdb/1.0.0/ Both are distributed by PhysioNet under the Open Data Commons Attribution License v1.0. code/01_download_gaitndd.py re-downloads them and verifies each file against the publisher's SHA-256 checksums, so every derived file here can be regenerated from the public sources. FIGURES The seven manuscript figures are not deposited as image files: they appear in the article, and code/03_figures.py rebuilds all of them from the tables in this deposit. REPRODUCING Scripts run under Python 3.11 (numpy, pandas, scipy, scikit-learn, xgboost, torch, shap, statsmodels, wfdb). Random seeds are fixed and every cross-validation split is subject-disjoint, so the reported figures reproduce to the digit. See README.md in the archive for the folder layout the scripts expect. DE-IDENTIFICATION Both source databases are distributed fully de-identified and no new human data were collected. Subjects appear only under the record identifiers assigned in the original PhysioNet releases; the only participant attributes present are those already published in the source databases.