SafeCKD-Agent: Code, Data, and Reproducibility Artifacts for Retrieval-Grounded Clinical Decision Support
Annotatsiya
SafeCKD-Agent: Code, Data, and Reproducibility Artifacts This Zenodo record contains the reproducibility package for the study: SafeCKD-Agent: A Trustworthy Agentic AI Framework for Retrieval-Grounded Clinical Decision Support The package provides code notebooks, public data artifacts, processed datasets, retrieval knowledge base files, trained model artifacts, prediction outputs, agentic workflow outputs, audit logs, evaluation tables, generated figures, manuscript-ready materials, metadata, and environment files. Contents of this Record This deposit includes: Source code notebooks for data exploration and result inspection. Public CKD and NHANES 2017–2018 source data artifacts used in the experiments. Cleaned, encoded, and split datasets for model evaluation. CKD retrieval knowledge base used by the SafeCKD-Agent workflow. Trained machine-learning model artifacts. Baseline prediction outputs. Agentic decision-support outputs. Agent audit logs. Evaluation tables and metrics. Generated figures and paper-ready figures. Case-study outputs. Reproducibility statements and safety statements. Directory tree, file manifest, checksums, and citation metadata. Python environment files for reproducible artifact inspection. Main Experimental Components Covered The deposited artifacts support inspection of the following components: UCI CKD benchmark experiments. NHANES 2017–2018 external validation. Conservative NHANES feature-sensitivity analysis. Baseline model comparison. Cross-validation results. Confusion-matrix and class-wise evaluation. Retrieval-quality evaluation. Agentic ablation analysis. Trustworthiness metrics. Safety stress testing. Hallucination/safety compliance outputs. Uncertainty-aware clinician-routing analysis. Calibration-related outputs. Runtime and auditability outputs. Paper-ready tables and figures. Directory Structure The package is organized as follows: SafeCKD/ │ ├── README.md ├── ZENODO_DESCRIPTION.md ├── CITATION.cff ├── LICENSE.md ├── DATASET_PROVENANCE.md ├── ETHICS_AND_LIMITATIONS.md ├── REPRODUCIBILITY.md ├── ENVIRONMENT.md ├── DIRECTORY_TREE.md ├── TODO_BEFORE_ZENODO.md ├── requirements.txt ├── environment.yml ├── Dockerfile │ ├── code/ │ ├── notebooks/ │ │ ├── 01_exploration.ipynb │ │ └── 02_results_analysis.ipynb │ └── scripts/ │ ├── data/ │ ├── raw/ │ │ ├── chronic_kidney_disease_ucimlrepo.csv │ │ ├── chronic_kidney_disease.zip │ │ ├── Chronic_Kidney_Disease.rar │ │ └── nhanes/ │ │ └── 2017-2018/ │ │ ├── ALB_CR_J.XPT │ │ ├── BIOPRO_J.XPT │ │ ├── BMX_J.XPT │ │ ├── BPQ_J.XPT │ │ ├── BPX_J.XPT │ │ ├── DEMO_J.XPT │ │ ├── DIQ_J.XPT │ │ ├── GHB_J.XPT │ │ └── GLU_J.XPT │ │ │ ├── processed/ │ │ ├── full_cleaned.csv │ │ ├── full_encoded.csv │ │ ├── train.csv │ │ ├── val.csv │ │ ├── test.csv │ │ ├── train_encoded.csv │ │ ├── val_encoded.csv │ │ ├── test_encoded.csv │ │ ├── nhanes_cleaned.csv │ │ ├── nhanes_train.csv │ │ ├── nhanes_val.csv │ │ └── nhanes_test.csv │ │ │ └── knowledge_base/ │ └── ckd_knowledge_base.json │ ├── outputs/ │ ├── models/ │ ├── predictions/ │ ├── agent_logs/ │ ├── tables/ │ ├── figures/ │ └── case_studies/ │ ├── paper_ready/ │ ├── figures/ │ ├── tables/ │ └── text/ │ ├── metadata/ │ ├── file_manifest.csv │ ├── checksums_sha256.txt │ └── data_files_overview.csv │ └── archives/ ├── outputs.zip └── paper_ready.zip Data Included Raw public data artifacts The data/raw/ folder contains public source data artifacts used in the experiments, including: UCI CKD dataset files. NHANES 2017–2018 .XPT files. Laboratory, examination, demographic, and questionnaire files used for external validation and sensitivity analysis. Processed data The data/processed/ folder contains: Cleaned datasets. Encoded datasets. Train/validation/test splits. UCI-style processed files. NHANES processed files. Knowledge base The data/knowledge_base/ folder contains: ckd_knowledge_base.json This file is used by the retrieval component of the SafeCKD-Agent workflow. Output Files The outputs/ folder contains generated artifacts from the study workflow, including: Model artifacts Trained baseline models. Preprocessor objects. Best-model information files. UCI and NHANES model artifacts. Predictions Baseline predictions. NHANES baseline predictions. Agentic workflow outputs. Single-step reasoning outputs. Agent logs Agentic audit logs. NHANES agentic audit logs. Evaluation tables Baseline metrics. Cross-validation summaries. External validation results. Conservative NHANES results. Agentic ablation results. Trustworthiness metrics. Retrieval-quality metrics. Safety stress-test results. Clinician-routing metrics. Calibration and uncertainty metrics. Confusion-matrix class-wise metrics. Statistical-test outputs. Figures Confusion matrices. ROC curves. Precision-recall curves. Feature-sensitivity plots. Agentic ablation figures. Retrieval-quality figures. Safety-compliance figures. Human-oversight figures. Calibration curves. Model-comparison plots. Paper-ready workflow and result figures. Paper-Ready Materials The paper_ready/ folder contains manuscript-ready outputs, including: Final figures used in the paper. Final tables used in the paper. Reproducibility checklist. Reproducibility statement. Safety statement. Limitations text. Results summary. Metadata and Integrity Files The metadata/ folder contains: file_manifest.csv checksums_sha256.txt data_files_overview.csv These files are included for file tracking and integrity verification. Users can verify downloaded files using the SHA256 checksums. Python Requirements The package includes a requirements.txt file for Python/pip-based setup. Main Python dependencies include: pandas>=2.0 numpy>=1.24 scipy>=1.10 scikit-learn>=1.3 xgboost>=2.0 joblib>=1.3 matplotlib>=3.7 seaborn>=0.13 shap>=0.44 openpyxl>=3.1 jupyterlab>=4.0 notebook>=7.0 pyreadstat>=1.2 Recommended Computational Environment Recommended environment: Python: 3.10 Operating system: Windows, Linux, or macOS Main libraries: pandas NumPy SciPy scikit-learn XGBoost SHAP joblib Matplotlib Seaborn OpenPyXL PyReadStat JupyterLab / Notebook Installation Options Option 1: Python venv + pip Windows: python -m venv .venv .venv\Scripts\activate python -m pip install --upgrade pip pip install -r requirements.txt jupyter lab Linux/macOS: python -m venv .venv source .venv/bin/activate python -m pip install --upgrade pip pip install -r requirements.txt jupyter lab Option 2: Conda/Mamba Using mamba: mamba env create -f environment.yml mamba activate safeckd-agent jupyter lab Using conda: conda env create -f environment.yml conda activate safeckd-agent jupyter lab Option 3: Docker A minimal Dockerfile is included for containerized artifact inspection: docker build -t safeckd-agent:1.0.0 . docker run --rm -p 8888:8888 -v "$PWD:/workspace" safeckd-agent:1.0.0 Then open the Jupyter URL shown in the terminal. Suggested Artifact Inspection Workflow After downloading and extracting the package: Open README.md for package overview. Check DIRECTORY_TREE.md for the full file structure. Check DATASET_PROVENANCE.md for data-source and processing notes. Create the Python environment using requirements.txt, environment.yml, or Docker. Open code/notebooks/01_exploration.ipynb for data inspection. Open code/notebooks/02_results_analysis.ipynb for output inspection. Inspect outputs/tables/ for reported metrics. Inspect outputs/figures/ and paper_ready/figures/ for generated figures. Inspect outputs/agent_logs/ for agentic audit traces. Use metadata/checksums_sha256.txt to verify file integrity. Reproducibility Level This package supports: Inspection of raw public data artifacts. Inspection of processed datasets. Inspection of trained model artifacts. Inspection of prediction outputs. Inspection of agentic workflow outputs. Inspection of evaluation tables and generated figures. Verification of paper-ready results. Environment recreation for notebook-based artifact inspection. SHA256-based file integrity checking. Full end-to-end regeneration from raw data may require adding the complete pipeline scripts if they exist outside this deposited package. The notebooks mention a run_all.py pipeline runner, but this file was not present during packaging. License and Third-Party Data Terms Recommended license for author-created materials: CC BY 4.0 for documentation, figures, tables, notebooks, metadata, and derived artifacts. Third-party public datasets and source archives retain their original licenses, citation requirements, and usage terms. Users should verify the terms of the original data providers before reuse. Ethics and Limitations This package is intended for: Research reproducibility. Auditability. Non-clinical research reuse. Independent inspection of reported results. This package is not intended for: Clinical deployment. Patient-specific diagnosis. Patient triage. Patient monitoring. Treatment recommendation. Use as a standalone medical device. SafeCKD-Agent and the included artifacts require clinical expert review before any translational or clinical use. Any downstream clinical use would require prospective validation, institutional approval, clinical governance, and complia
Hali tarjima qilinmagan