Vivli - Anticipating Resistance Risks to Cefiderocol in MDR Pathogens

Overview

This project combines multiple prediction models to provide evidence-based antibiotic recommendations. The system uses machine learning to recommend antibiotics in optimal order of efficacy, with cefiderocol as the last resort option.

🏗️ Architecture

The system follows a structured 4-step methodology:

1. Step 1: Data Preparation and Exploration
2. Step 2: Antibiotic Decision Tree Model Development
3. Step 3: Phenotypic Signature Analysis and Clustering
4. Step 4: Cefiderocol Use Prediction Model

📁 Project Structure

Vivli/
├── scripts/                    # Python and R scripts
│   ├── antibiotic_decision_tree.py
│   ├── step4_prediction.py
│   ├── generate_english_antibiotic_report.py
│   ├── convert_md_to_html.py
│   ├── multiple_regression_plot.R
│   ├── univariate_analysis_script.R
│   └── ...
├── docs/                       # Documentation
│   ├── vivli_complete_methodology.md
│   ├── vivli_complete_methodology.html
│   ├── antibiotic_recommendation_system_details_english.md
│   ├── last_prediction_model_details.md
│   └── ...
├── outputs/                    # Generated outputs
│   ├── reports/               # PDF and HTML reports
│   ├── plots/                 # Visualizations
│   └── models/                # Trained models
├── data/                      # Data files (not included in repo)
│   ├── 1.xlsx                # SIDERO-WT Database
│   └── 2.xlsx                # ATLAS Database
└── README.md

🚀 Quick Start

Prerequisites

pip install pandas numpy scikit-learn matplotlib seaborn xgboost shap reportlab markdown

Running the System

1. Antibiotic Decision Tree Model:

cd scripts
python antibiotic_decision_tree.py

2. Cefiderocol Prediction Model:

cd scripts
python step4_prediction.py

3. Generate English Report:

cd scripts
python generate_english_antibiotic_report.py

📊 Models

1. Antibiotic Decision Tree Model

• Algorithm: Decision Tree Classifier
• Features: 7 primary features (species, country, year, resistance patterns)
• Performance: 100% accuracy (reported)
• Output: Complete antibiotic sequence with cefiderocol as last resort

2. Cefiderocol Use Prediction Model

• Algorithm: Random Forest Classifier
• Features: 20+ features including MIC values, resistance patterns, ratios
• Performance: AUC 1.000, Precision 1.000, Recall 1.000
• Output: Binary decision for cefiderocol use

3. Phenotypic Signature Analysis

• Method: Clustering analysis with PCA
• Purpose: Identify resistance patterns and signatures
• Output: Cluster assignments and phenotypic signatures

📈 Performance Metrics

Antibiotic Decision Tree

• Accuracy: 100%
• Target: First antibiotic recommendation
• Coverage: 40+ antibiotics analyzed

🏥 Clinical Applications

Decision Framework

1. First-Line Treatment: Most effective antibiotic based on species, region, resistance
2. Sequential Alternatives: Complete sequence of alternatives
3. Phenotypic Analysis: Resistance patterns and clusters
4. Last Resort Decision: Cefiderocol use determination

Example Usage

# Get antibiotic recommendations
recommendations = model.recommend_antibiotics(
    species="Escherichia coli",
    country="France", 
    year=2023,
    resistance_profile={'beta_lactam': 0.3, 'quinolone': 0.7}
)

📋 Data Sources

• ATLAS Database (2.xlsx): Global antimicrobial susceptibility data

  • 966,805 isolates, 134 variables
  • Multiple countries and species
  • Temporal coverage

• SIDERO-WT Database (1.xlsx): Cefiderocol-specific susceptibility data

  • MIC values and resistance patterns
  • Species and geographic information

⚠️ Important Limitations

Critical Considerations

• No real treatment failure data available
• Targets based on theoretical resistance patterns
• High performance likely reflects simplified target definitions
• Geographic and temporal biases possible

Clinical Implementation

• Do NOT use for clinical decisions without validation
• Validate on real treatment failure data
• Include clinical factors (comorbidities, previous exposure)
• Prospective validation required before implementation
• Use as research tool only

🔬 Technical Details

Feature Engineering

• MIC value standardization
• Resistance threshold application
• Categorical encoding
• Composite resistance scores
• MIC ratios for comparative analysis

Model Training

• Train/Test Split: 80/20
• Cross-validation: 5-fold StratifiedKFold
• Feature scaling: StandardScaler
• Random State: 42

Dependencies

• Scikit-learn: Machine learning algorithms
• XGBoost: Gradient boosting
• SHAP: Feature importance analysis
• Pandas/NumPy: Data manipulation
• Matplotlib/Seaborn: Visualization
• ReportLab: PDF generation

📚 Documentation

Complete Documentation

• Methodology: docs/vivli_complete_methodology.html
• Antibiotic System: docs/antibiotic_recommendation_system_details_english.html
• Cefiderocol Model: docs/last_prediction_model_details.html

Reports

• English PDF Report: outputs/reports/antibiotic_recommendation_report_english.pdf
• Methodology HTML: docs/vivli_complete_methodology.html

🚀 Future Directions

1. Clinical Integration

• Include patient factors (age, comorbidities, allergies)
• Add pharmacokinetic considerations
• Integrate with local resistance patterns

2. Model Improvements

• Include genomic resistance markers
• Add temporal resistance trends
• Develop species-specific models

3. Clinical Validation

• Prospective clinical studies
• Real-world implementation
• Outcome assessment

4. Broader Applications

• Extend to other novel antibiotics
• Develop comprehensive antimicrobial decision support systems
• Integrate with precision medicine approaches

Authors

Adekemi Adepeju, Christian Ako, Abeeb Adeniyi, Oluwatobiloba Kazeem, Oluwadamilare Olatunbosun using the Pfizer's ATLAS dataset and Sidero datatset as part of the 2025 Vivli AMR Data Challenge.