MLPA Digital Twin Survival Analysis for NSCLC

A comprehensive survival analysis framework integrating **Multi-Level Parameterized Automata (MLPA) ** simulations with radiomics, clinical, and deep learning features for Non-Small Cell Lung Cancer (NSCLC) prognosis prediction.

** Try the Interactive Demo** - Visualize 2-D tumor growth simulation in your browser!

---

🎯 Overview

This implementation replicates Gu J et al. (2025) methodology and extends it with:

• MLPA Integration: 3-D cellular automaton simulations of tumor growth
• Multiple Models: CoxPH, Neural Cox (DeepSurv), Gradient Boosting, Ensemble
• Integrated AUC (iAUC): Time-dependent discrimination assessment
• Sensitivity Analysis: Parameter robustness testing

Pipeline:

CT Scans → Radiomics → Biological Parameters (α, β) → 3D Cellular Automaton → Growth Features

---

✨ Features

• 🔬 4 Survival Models: CoxPH, Neural Cox, Gradient Boosting, Ensemble
• 📊 4 Metrics: C-Index, p-value, Hazard Ratio, iAUC
• 🧬 4 Data Domains: Clinical (7), Radiomics (1706), AutoEncoder (512), Digital Twin (6)
• 🔀 2 Fusion Methods: Feature-level (Table 5), Signature-level (Table 6)
• 🔍 Sensitivity Analysis: α/β parameter perturbation (±10%, ±20%)
• 📈 Kaplan-Meier Curves: Risk stratification visualization

---

📋 Quick Start

Installation

# Core dependencies (required)
pip install numpy pandas scikit-learn matplotlib lifelines tqdm

# Optional (for full functionality)
pip install torch scikit-survival  # Neural Cox + iAUC

Data Structure

project_root/
├── data3d_1706features/outputs/
│   ├── features_normalized_standard.npy
│   ├── feature_names_normalized_standard.json
│   └── patient_ids_normalized_standard.json
├── nsclc/NSCLC-Radiomics-Lung1.clinical-version3-Oct-2019.csv
├── data_3d_ae_parallel/features/features_deep_512.csv
└── digital_twin_output/all_patients_biological_parameters.csv

Run Analysis

python ferretti_3d_mlpa_iauc_neural_ensemble.py

Runtime: 10-20 minutes (with Neural Cox), 2-5 minutes (CoxPH only)

---

📊 Key Configuration

Edit at top of script:

# Feature selection
TOP_K_AE = 15
TOP_K_RADIOMICS = 10
CORRELATION_THRESHOLD = 0.70

# Cross-validation
N_SPLITS = 5
RANDOM_STATE = 42

# Neural Cox
NEURAL_COX_HIDDEN_LAYERS = [64, 32]
NEURAL_COX_DROPOUT = 0.3
NEURAL_COX_EPOCHS = 100

# Ensemble weights
ENSEMBLE_WEIGHTS = {
    'coxph': 0.4,
    'neural_cox': 0.3,
    'gradient_boosting': 0.3
}

# Model selection
RUN_COXPH = True
RUN_NEURAL_COX = True
RUN_ENSEMBLE = True
CREATE_KM_PLOTS = True

# iAUC time points (years)
IAUC_TIME_POINTS = [0.25, 0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0]

---

📁 Outputs

All saved to ./ferretti_with_3d_mlpa_iauc_output/:

File	Description
results_with_3d_mlpa_iauc_multimodel.csv	Complete results (C-Index, HR, iAUC)
results_pivot_by_model.csv	Model comparison table
results_sorted_by_iauc.csv	Ranked by integrated AUC
km_plots/*.png	Kaplan-Meier curves (high vs low risk)
sensitivity_analysis/sensitivity_summary.csv	Parameter perturbation results
sensitivity_analysis/*.png	Robustness visualizations

---

📈 Example Results

┌─────────────────────────────────┬──────────┬────────────────┬────────────┬────────────┬────────────────┐
│ Configuration                   │ Model    │ C-Index        │ p-value    │ HR         │ iAUC           │
├─────────────────────────────────┼──────────┼────────────────┼────────────┼────────────┼────────────────┤
│ Table5_DT_R_C                   │ ensemble │ 0.6523±0.0287  │  0.0012    │ 2.84±0.45  │ 0.6891±0.0312  │
│ Table5_R                        │ neural   │ 0.6401±0.0295  │  0.0024    │ 2.67±0.38  │ 0.6734±0.0289  │
│ Table6_C                        │ coxph    │ 0.6278±0.0301  │  0.0031    │ 2.54±0.41  │ 0.6612±0.0276  │
└─────────────────────────────────┴──────────┴────────────────┴────────────┴────────────┴────────────────┘

Interpretation:

• C-Index > 0.7: Excellent | 0.6-0.7: Good | 0.5-0.6: Moderate
• iAUC > 0.7: Excellent | 0.6-0.7: Good | 0.5-0.6: Moderate
• p-value < 0.05: Significant | HR > 2.0: Strong risk stratification

---

🧠 Features

The 3-D MLPA generates 6 features from radiomics:

Feature	Description	Derivation
Bio_Alpha	Proliferation rate (α)	0.01 + 0.05 × entropy
Bio_Necrosis	Necrosis probability (β)	0.01 + 0.08 × (1 - sphericity)
Sim_Growth_Rate	Simulated expansion rate	3D cellular automaton
Sim_Necrosis_Ratio	Necrotic cell fraction	3D cellular automaton
Radiomics_Entropy	Tumor heterogeneity	First-order histogram
Radiomics_Sphericity	Shape regularity	Surface area / volume

Why MLPA?

• ✅ Biological interpretability (α, β have clinical meaning)
• ✅ Physics-based constraints (realistic growth dynamics)
• ✅ Patient-specific personalization
• ✅ Longitudinal prediction capability

---

🔬 Model Comparison

Model	Pros	Cons	Use Case
CoxPH	Fast, interpretable, established	Linear assumptions	Baseline, coefficient interpretation
Neural Cox	Non-linear, flexible	Black box, slower	Complex interactions
Gradient Boosting	Robust, handles interactions	Less interpretable	Tree-based ensemble
Ensemble	Best of all, reduces variance	Slower, more complex	Best performance

---

🔍 Sensitivity Analysis

Tests model robustness by perturbing α and β by ±10%, ±20%:

Metrics:

• Risk score changes (absolute & percentage)
• Rank preservation (Spearman ρ)
• Risk category changes (% switching high/low)
• C-Index stability (ΔC-Index)

Robust model criteria:

• Spearman ρ > 0.95
• Category changes < 10%
• |ΔC-Index| < 0.02

---

🐛 Troubleshooting

Issue	Solution
"scikit-survival not installed"	pip install scikit-survival (enables iAUC)
"PyTorch not installed"	pip install torch (enables Neural Cox)
"Digital Twin file not found"	Script continues without DT features
Low C-Index (<0.55)	Increase TOP_K_AE, TOP_K_RADIOMICS
Out of memory	Reduce NEURAL_COX_BATCH_SIZE to 16
Slow execution	Set RUN_NEURAL_COX=False, N_SPLITS=3

---

📚 Citation

@software{phong2026mlpa,
  author       = {Huu Phong Nguyen, Delower Hossain, Ehsan Saghapour, Zhandos Sembay, Jake Y. Chen 
},
  title        = {{MLPA: A Framework Modeling for Non-Small Cell Lung Cancer Survival Prediction 
}},
  year         = {2026},
  institution  = {Systems Pharmacology AI Research Center (SPARC), 
                  University of Alabama at Birmingham},
  type         = {Software Framework},
  url          = {[https://github.com/aimed-lab/MLPA_abstract](https://github.com/aimed-lab/MLPA_abstract)},
  note         = {3-D Multi-Level Parameterized Automata for NSCLC survival prediction}
}

---

📧 Contact

Author: Huu Phong Nguyen (hnguye24 AT uab DOT edu)

Affiliation: SPARC (Systems Pharmacology AI Research Center), UAB

Project: Lung Cancer MLPA Framework

---

⚖️ License

MIT License - See LICENSE file for details

---

Built with ❤️ for advancing precision oncology through mechanistic AI