SynDX: Explainable AI-Guided Synthetic Data Generation Framework

PhD Research Project - XAI-Guided Parameter Space Exploration for Clinical Synthetic Data Generation

Overview

This repository contains the implementation of a novel framework for generating clinically valid synthetic medical data using explainable AI (XAI) techniques. The framework integrates three XAI methods—SHAP, Counterfactual Explanations, and Non-negative Matrix Factorization (NMF)—to guide parameter space exploration and ensure clinical validity, interpretability, and epidemiological fidelity.

Domain: Vestibular disorders (dizziness and vertigo) Target Journal: Nature Machine Intelligence Status: Research implementation for PhD dissertation

Key Innovations

1. XAI-Guided Generation: First framework to integrate explainability methods (SHAP, Counterfactuals, NMF) during synthetic data generation, not just after
2. Clinical Constraint Enforcement: TiTrATE framework with 10 Boolean constraints ensures 100% clinically valid patients
3. Multi-Method Validation: Comprehensive empirical validation with 319+ metrics across 4 XAI domains
4. Zero Constraint Violations: Achieves 0% clinical constraint violations vs 16-23% for state-of-the-art GANs

Mathematical Framework

The framework implements a 6-phase exploration algorithm guided by three XAI methods:

• Phase 1: Uniform sampling from parameter space
• Phase 2: NMF factor discovery (Eq. 15-18)
• Phase 3: SHAP feature importance analysis (Eq. 19-21)
• Phase 4: Importance-weighted sampling (60% budget)
• Phase 5: Critical scenario targeting (15% coverage)
• Phase 6: Diversity-aware sampling (DPP-based)

Complete implementation of all formulas from manuscript:

• Eq. 1-8: Parameter space definitions
• Eq. 9-14: Target archetype calculation
• Eq. 15-18: NMF decomposition
• Eq. 19-21: SHAP importance weighting
• Eq. 22-24: Epidemiological distributions
• Algorithm 7.1: Complete 6-phase exploration pipeline

Installation

# Clone repository
git clone https://github.com/ChatchaiTritham/SynDX.git
cd SynDX

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Quick Start

from syndx.phase1_knowledge.domain_config import create_vestibular_domain
from syndx.phase1_knowledge.xai_explorer import XAIGuidedExplorer

# Create parameter space (126,000 combinations)
param_space = create_vestibular_domain()

# Initialize XAI-guided explorer
explorer = XAIGuidedExplorer(
    parameter_space=param_space,
    n_target=8400,
    nmf_factors=10,  # 10 latent clinical factors
    random_state=42
)

# Execute 6-phase exploration
archetypes = explorer.explore()

print(f"Generated {len(archetypes)} clinically valid archetypes")

Repository Structure

SynDX/
├── syndx/               # Core framework implementation
│   ├── core/           # Parameter space, constraints, epidemiology
│   ├── phase0_uniform/ # Uniform sampling
│   ├── phase1_knowledge/ # NMF discovery, SHAP analysis
│   ├── phase2_synthesis/ # Importance sampling, critical targeting
│   ├── phase3_validation/ # Validation metrics
│   └── utils/          # Helper functions
├── examples/           # Usage examples and demos
├── tests/             # Unit tests
├── data/              # Domain configuration files
└── docs/              # Documentation

Key Files:
- VALIDATION_PROTOCOL.md: Complete empirical validation methodology
- PUBLICATION_CHECKLIST.md: Manuscript preparation checklist
- EXPERT_EVALUATION_TEMPLATE.md: Clinical expert evaluation protocol

Key Features

1. Clinical Validity Enforcement

TiTrATE Constraint Framework ensures all generated patients are clinically plausible:

• 10 Boolean constraints covering age-diagnosis relationships, symptom consistency, temporal patterns
• 100% constraint satisfaction (vs 16-23% violations for MedGAN/CTGAN/VAE)
• Expert validation: 91.3% agreement (Cohen's κ = 0.89)

2. Explainability Integration

Three XAI methods guide generation:

• SHAP: Identifies most important features for importance-weighted sampling
• Counterfactuals: Generates minimal-change diagnostic transitions (70% success, 3.31 features changed, 85% clinical plausibility)
• NMF: Discovers 10 latent clinical factors with 46-56% sparsity

3. Comprehensive Validation

319+ validation metrics across multiple dimensions:

• SHAP stability: Kendall τ ≈ 0 (near-perfect rank consistency), bootstrap ρ = 0.67
• Counterfactual quality: 70% success rate, mean sparsity 3.31, plausibility 3.78/5
• NMF interpretability: 10 factors, ANOVA p < 0.001 for disease associations
• Framework performance: 0.7% utility gap (best-in-class vs baselines)

4. State-of-the-Art Performance

Comparison against 4 baselines (MedGAN, CTGAN, Synthea, VAE):

• Best diagnostic utility: 93.5% accuracy (0.7% gap from real data)
• Best XAI fidelity: SHAP ρ = 0.89 (vs 0.61-0.74 for baselines)
• Best statistical parity: KL divergence = 0.12 (vs 0.21-0.35)
• Zero constraint violations: 0% (vs 16-23% for GANs)
• No mode collapse: 97.3% rare diagnosis coverage (vs 26% for MedGAN)

Results

Expected output for vestibular domain (15 diagnoses):

• Parameter space: 126,000 total combinations
• Valid archetypes: 60,480 (48% acceptance rate)
• Generated: 8,400 archetypes
• NMF factors: r = 10 latent clinical phenotypes
• Critical coverage: 15.0% (stroke/TIA, perfect match to target)
• Epidemiological fidelity: χ² p = 0.18 (not significantly different from real population)
• Computational cost: 27 minutes on 16-core CPU, 3.5 GB memory (laptop-friendly)

Validation

Complete empirical validation following Nature Machine Intelligence standards:

• Bootstrap resampling (n=1,000 iterations) for all confidence intervals
• Expert clinical evaluation (3 clinicians, Krippendorff's α = 0.72)
• Statistical significance testing (McNemar's, ANOVA, χ², Bonferroni-corrected)
• Critical diagnosis analysis (stroke/TIA sensitivity 93.3%)
• Rare diagnosis coverage (all ≥50 examples, no mode collapse)

See VALIDATION_PROTOCOL.md for complete methodology.

Publications

Manuscript in Preparation:

• Title: "SynDX: Explainable AI-Guided Synthetic Data Generation for Clinical Domains"
• Target: Nature Machine Intelligence
• Status: 85% complete (main text, supplementary materials ready)

Citation (preprint):

@article{syndx2026,
  title={SynDX: Explainable AI-Guided Parameter Space Exploration for Clinical Synthetic Data Generation},
  author={Chatchai Tritham},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2026},
  note={PhD Dissertation, Naresuan University}
}

Requirements

• Python 3.8+
• NumPy, SciPy, Pandas
• scikit-learn 1.3+
• XGBoost 2.0+
• SHAP 0.42+
• Matplotlib, Seaborn (for visualization)

See requirements.txt for complete dependencies.

Reproducibility

All experiments are fully reproducible:

• Random seed: 42 (set throughout codebase)
• Software versions documented in requirements.txt
• Hardware specifications: 16-core CPU, 32GB RAM, optional GPU
• Complete hyperparameter specifications in manuscript

License

MIT License - see LICENSE file

Contact

Author

Chatchai Tritham (PhD Candidate)

• Email: chatchait66@nu.ac.th
• Department of Computer Science and Information Technology
• Faculty of Science, Naresuan University
• Phitsanulok 65000, Thailand

Supervisor

Chakkrit Snae Namahoot

• Email: chakkrits@nu.ac.th
• Department of Computer Science and Information Technology
• Faculty of Science, Naresuan University
• Phitsanulok 65000, Thailand

For Questions or Collaborations

• GitHub Issues: https://github.com/ChatchaiTritham/SynDX/issues
• Email: chatchait66@nu.ac.th

Acknowledgments

This research is part of a PhD dissertation at Naresuan University. The framework was developed to address the critical need for interpretable and clinically valid synthetic medical data generation, with applications in privacy-preserving machine learning research and rare disease studies.

Ethical Statement: This research uses only synthetic data generated from a rule-based framework. No real patient data was used. IRB approval was not required.

---

Last Updated: January 2026 Repository Status: Active Development Manuscript Status: In Preparation for Nature Machine Intelligence