🚶🏻♀️ PB2PD — Wearable Gait Biomarkers and Explainable AI for High Retrospective Prodromal Burden in Parkinson’s Disease
Wearable Gait Biomarkers and Explainable AI Identify High Retrospective Prodromal Burden in Parkinson’s Disease.
The project investigates whether wearable-derived gait features are associated with high retrospective prodromal/non-motor burden within established Parkinson’s disease (PD). The analyses use lumbar inertial-sensor gait features, clinical covariates, machine-learning models, and explainability methods to characterize a within-PD digital phenotype.
All participants included in the study had established Parkinson’s disease at the time of gait assessment. Therefore, the PB+ label used in this project should be interpreted as a marker of high retrospective prodromal/non-motor burden within diagnosed PD, not as formal prodromal PD status and not as a pre-diagnostic classification rule.
The analyses do not establish prospective conversion risk, early detection performance, or diagnostic validity in undiagnosed or at-risk populations. Future longitudinal studies in pre-diagnostic cohorts are required to determine whether similar gait signatures are detectable before clinical PD diagnosis.
This repository contains the notebook-based analysis pipeline associated with the manuscript:
Wearable Gait Biomarkers and Explainable AI Identify High Retrospective Prodromal Burden in Parkinson’s Disease
The repository is intended to support transparency, reproducibility, and peer review of the computational workflow reported in the manuscript.
This repository contains the ordered notebook workflow, software requirements, and documentation needed to understand the analysis sequence.
PB2PD/
├── README.md
├── requirements.txt
└── notebooks/
├── 00_target_definition_supporting_analysis.ipynb
├── 01_eda.ipynb
├── 02_preprocessing.ipynb
├── 03_balancing.ipynb
├── 04_model_explainability.ipynb
├── 05_subgroup_robustness_and_clinical_appendix.ipynb
├── 06_leakage_safe_pipeline_check.ipynb
└── 07_reviewer_response_additional_analyses.ipynb
The notebooks are numbered according to the intended execution order.
00_target_definition_supporting_analysis.ipynb
Exploratory target-definition analysis for retrospective prodromal/non-motor burden formulations
01_eda.ipynb
Exploratory data analysis and descriptive/statistical characterization
02_preprocessing.ipynb
Data cleaning, preprocessing, missing-data handling, and feature-selection workflow
03_balancing.ipynb
Class-imbalance management, CTGAN-based training-set augmentation, and model-development analyses
04_model_explainability.ipynb
Random Forest modelling, SHAP explainability, surrogate rules, and interaction analyses
05_subgroup_robustness_and_clinical_appendix.ipynb
Subgroup, calibration, robustness, and clinical appendix analyses
06_leakage_safe_pipeline_check.ipynb
Reviewer-response leakage-safety checks, including train-only preprocessing/augmentation and held-out real test evaluation
07_reviewer_response_additional_analyses.ipynb
Additional reviewer-response analyses, including prodromal symptom prevalence, exploratory clustering, paired bootstrap comparisons, fold-wise performance variability, and revised CTGAN validation summaries
The computational workflow includes:
- definition and characterization of high retrospective prodromal/non-motor burden;
- descriptive and exploratory statistical analyses;
- preprocessing, missing-data handling, and feature selection;
- Random Forest classification;
- CTGAN-based training-set augmentation for class-imbalance management;
- held-out real test-set evaluation;
- bootstrap-based uncertainty estimation;
- SHAP-based global and local explainability;
- surrogate-rule extraction;
- subgroup and robustness analyses;
- reviewer-response sensitivity analyses.
CTGAN-generated samples were used only as a training-support strategy for class-imbalance management. They should not be interpreted as biologically validated synthetic patient profiles. All clinical and biomechanical interpretations in the manuscript are based on real observations and explainability analyses of the trained model.
The clinical dataset is not included directly in this code repository because it contains human-subject clinical information. Data availability is described in the manuscript.
Users with approved access to the de-identified analysis-ready dataset should place the dataset in the expected data directory before running the notebooks.
Create a Python environment and install the required packages:
pip install -r requirements.txt
The analyses were developed using Python 3.11.
The analyses reported in the manuscript were performed on a MacBook Pro workstation equipped with an Apple M3 Pro processor, 18 GB RAM, and macOS Sequoia 15.6.1.
Random Forest modelling, cross-validation, bootstrap analyses, and SHAP explainability were performed on CPU. No dedicated GPU acceleration was required for the analyses reported in the study.
To reproduce the workflow:
- Clone the repository.
- Install the dependencies listed in requirements.txt.
- Obtain approved access to the de-identified analysis-ready dataset, if applicable.
- Place the dataset in the expected data/raw/ directory.
- Run the notebooks sequentially from 00 to 07.
Because notebook outputs are not stored in the repository, figures and tables will be regenerated when the notebooks are executed in the appropriate environment with access to the required data.
https://doi.org/10.5281/zenodo.17551802
Trabassi D. et al. Wearable Gait Biomarkers and Explainable AI Identify High Retrospective Prodromal Burden in Parkinson’s Disease.
This repository contains research code only. The analyses are intended for scientific reproducibility and peer review.
The model is not a certified medical device and is not intended for clinical diagnosis, prospective screening, individual-level risk prediction, or medical decision-making.