Implementation of the pipeline described in:
Bonazzetti, C., Rocchi, E., Toschi, A. et al. Artificial Intelligence model to predict resistances in Gram-negative bloodstream infections. npj Digit. Med. 8, 319 (2025). https://doi.org/10.1038/s41746-025-01696-x
Website | Documentation | Installation | Quick Start | Pipeline Overview | CLI Commands | Citation
A reproducible machine learning framework designed to accelerate clinical decision-making by predicting antimicrobial resistance patterns from patient data.
- Project Website - Overview of the project, original paper, and related work
- Documentation - Installation guides, CLI reference, tutorials, and API documentation
Install from PyPI:
pip install respredaiOr install from source:
git clone https://github.com/EttoreRocchi/ResPredAI.git
cd ResPredAI
# For development (includes pytest)
pip install -e ".[dev]"Verify the installation:
respredai --versionrespredai create-config my_config.iniEdit my_config.ini with your data paths and parameters:
[Data]
data_path = ./data/my_data.csv
targets = Target1,Target2
continuous_features = Feature1,Feature2,Feature3
[Metadata]
# group_column = PatientID # Optional: prevents data leakage in CV
# temporal_column = collection_date # Date column for temporal validation
# subgroup_columns = ward, sex # Columns for subgroup performance analysis
[Pipeline]
models = LR,RF,XGB,CatBoost
outer_folds = 5
inner_folds = 3
# Repeated CV: set >1 for more robust estimates
outer_cv_repeats = 1
# Probability calibration: post-hoc calibration on best estimator
calibrate_probabilities = false
probability_calibration_method = sigmoid # sigmoid or isotonic
probability_calibration_cv = 5
# Threshold optimization
calibrate_threshold = false
threshold_method = auto
# Threshold optimization objective: youden (default), f1, f2, cost_sensitive
threshold_objective = youden
# Cost weights for cost_sensitive objective (VME = false susceptible, ME = false resistant)
vme_cost = 1.0
me_cost = 1.0
# Confidence level for bootstrap CIs (between 0.5 and 1.0, default: 0.95)
confidence_level = 0.95
# Number of bootstrap resamples for CIs (>= 100, default: 1000)
n_bootstrap = 1000
[Uncertainty]
# Miscoverage rate for conformal prediction (default 0.1 = 90% coverage)
alpha = 0.1
[Reproducibility]
seed = 42
[Log]
verbosity = 1
log_basename = respredai.log
[Resources]
n_jobs = -1
[ModelSaving]
enable = true
compression = 3
[Preprocessing]
ohe_min_frequency = 0.05
[Imputation]
method = none
strategy = mean
n_neighbors = 5
estimator = bayesian_ridge
[Output]
out_folder = ./output/
[Validation]
# Validation strategy: cv (default), temporal (prospective-style), or both
validation_strategy = cv
# temporal_split_date = 2023-01-01 # Cutoff date (train < date, test >= date)
# temporal_split_ratio = 0.8 # Alternative: fraction for training (by date order)Tip: Comment out optional parameters with
#to disable them. Empty values (e.g.,group_column =) are treated as absent.
respredai run --config my_config.iniflowchart LR
A[Configuration Loading] --> B[Data Loading & Validation]
B --> C[OHE Template]
C --> D{Validation Strategy}
D -->|cv / both| E["Outer CV Loop (group-aware if configured)"]
D -->|temporal / both| F[Temporal Split]
E --> G["OHE (fit on train, transform test)"]
G --> H["Scaling (fit on train, transform test)"]
H --> I[Hyperparameter Tuning - Inner CV]
I --> J[Calibration & Threshold]
J --> K[Predict on Test Fold]
K --> L[Metrics + Bootstrap CIs]
L --> M[Subgroup Analysis]
F --> T1["OHE (fit on train, transform test)"]
T1 --> T2["Scaling (fit on train, transform test)"]
T2 --> T3[Hyperparameter Tuning - Inner CV]
T3 --> T6[Calibration & Threshold]
T6 --> T4[Predict on Test Split]
T4 --> T5[Metrics + Bootstrap CIs]
M --> R[Generate Reports]
T5 --> R
R --> R1[Summary CSVs + HTML Report]
R --> R2[Confusion Matrices]
R --> R3[Calibration Curves]
R --> R4[Feature Importance]
style A fill:#e8e8e8,stroke:#999,color:#333
style B fill:#e8e8e8,stroke:#999,color:#333
style C fill:#b2ebf2,stroke:#0dafb5,color:#333
style D fill:#ffe0b2,stroke:#f5a623,color:#333
style E fill:#b2ebf2,stroke:#0dafb5,color:#333
style F fill:#b2ebf2,stroke:#0dafb5,color:#333
style G fill:#b2ebf2,stroke:#0dafb5,color:#333
style H fill:#b2ebf2,stroke:#0dafb5,color:#333
style I fill:#b2ebf2,stroke:#0dafb5,color:#333
style J fill:#fff3e0,stroke:#f5a623,color:#333
style K fill:#b2ebf2,stroke:#0dafb5,color:#333
style L fill:#b2ebf2,stroke:#0dafb5,color:#333
style M fill:#fff3e0,stroke:#f5a623,color:#333
style T1 fill:#b2ebf2,stroke:#0dafb5,color:#333
style T2 fill:#b2ebf2,stroke:#0dafb5,color:#333
style T3 fill:#b2ebf2,stroke:#0dafb5,color:#333
style T6 fill:#fff3e0,stroke:#f5a623,color:#333
style T4 fill:#b2ebf2,stroke:#0dafb5,color:#333
style T5 fill:#b2ebf2,stroke:#0dafb5,color:#333
style R fill:#0dafb5,stroke:#098a8f,color:#fff
style R1 fill:#0dafb5,stroke:#098a8f,color:#fff
style R2 fill:#0dafb5,stroke:#098a8f,color:#fff
style R3 fill:#0dafb5,stroke:#098a8f,color:#fff
style R4 fill:#fff3e0,stroke:#f5a623,color:#333
flowchart LR
A[Configuration Loading] --> B[Data Loading & Validation]
B --> C[OHE on Full Data]
C --> E[Feature Scaling]
E --> F[Hyperparameter Tuning - Inner CV]
F --> G[Calibration & Threshold]
G --> H[Save Model Bundle]
H --> I["Output: model + transformer + OHE + threshold + metadata"]
style A fill:#e8e8e8,stroke:#999,color:#333
style B fill:#e8e8e8,stroke:#999,color:#333
style C fill:#b2ebf2,stroke:#0dafb5,color:#333
style E fill:#b2ebf2,stroke:#0dafb5,color:#333
style F fill:#b2ebf2,stroke:#0dafb5,color:#333
style G fill:#fff3e0,stroke:#f5a623,color:#333
style H fill:#0dafb5,stroke:#098a8f,color:#fff
style I fill:#0dafb5,stroke:#098a8f,color:#fff
flowchart LR
A[Load Training Metadata] --> B[Load New Data + Validate Features]
B --> D["OHE (fitted on training data, transform new data)"]
D --> E["Scaling (fitted on training data, transform new data)"]
E --> F[Predict with Saved Threshold]
F --> G[Metrics vs Ground Truth]
G --> H[Uncertainty Scores]
H --> I[Predictions + Metrics CSVs]
I --> J[Evaluation Summary]
style A fill:#e8e8e8,stroke:#999,color:#333
style B fill:#e8e8e8,stroke:#999,color:#333
style D fill:#b2ebf2,stroke:#0dafb5,color:#333
style E fill:#b2ebf2,stroke:#0dafb5,color:#333
style F fill:#b2ebf2,stroke:#0dafb5,color:#333
style G fill:#b2ebf2,stroke:#0dafb5,color:#333
style H fill:#fff3e0,stroke:#f5a623,color:#333
style I fill:#0dafb5,stroke:#098a8f,color:#fff
style J fill:#0dafb5,stroke:#098a8f,color:#fff
Amber nodes indicate optional steps controlled by configuration parameters. All pipelines are executed for each model × target combination.
respredai run --config path/to/config.ini [--quiet]Train models using nested cross-validation with the specified configuration.
📖 Detailed Documentation - Complete guide with all configuration options and workflow details.
respredai train --config path/to/config.ini [--models LR,RF] [--output ./trained/]Train models on the entire dataset using GridSearchCV for hyperparameter tuning. Saves one model file per model-target combination for later use with evaluate.
📖 Detailed Documentation - Complete guide with output structure and workflow.
respredai evaluate --models-dir ./output/trained_models --data new_data.csv --output ./eval/Apply trained models to new data with ground truth. Outputs predictions and metrics.
📖 Detailed Documentation - Complete guide with data requirements and output format.
respredai feature-importance --output <output_folder> --model <model_name> --target <target_name> [--top-n 20]Extract and visualize feature importance/coefficients from trained models across all outer cross-validation iterations. Uses SHAP as fallback for models without native feature importance.
📖 Detailed Documentation - Complete guide with interpretation, examples, and statistical considerations.
respredai list-modelsDisplay all available machine learning models with descriptions.
Available Models:
┌────────────┬──────────────────────────┐
│ Code │ Name │
├────────────┼──────────────────────────┤
│ LR │ Logistic Regression │
│ MLP │ Multi-Layer Perceptron │
│ XGB │ XGBoost │
│ RF │ Random Forest │
│ CatBoost │ CatBoost │
│ TabPFN │ TabPFN │
│ RBF_SVC │ RBF SVM │
│ Linear_SVC │ Linear SVM │
│ KNN │ K-Nearest Neighbors │
└────────────┴──────────────────────────┘
respredai create-config output_path.iniGenerate a template configuration file that you can edit for your data.
📖 Detailed Documentation - Complete guide to configuration file structure and customization.
respredai validate-config <path_to_config.ini> [--check-data]Validate a configuration file without running the pipeline. It can also check that the dataset load without errors.
📖 Detailed Documentation - Complete guide to configuration file validation.
respredai infoDisplay information about ResPredAI including scientific paper citation and version details.
Or just:
respredai --versionto show the installed version of ResPredAI.
The pipeline generates:
- Confusion matrices: PNG files with heatmaps showing model performance for each target
- Detailed metrics tables: CSV files with comprehensive metrics (precision, recall, F1, MCC, balanced accuracy, AUROC, VME, ME, Brier Score, ECE, MCE) with mean, std, and 95% CI
- Calibration diagnostics: Reliability curves (calibration plots) per fold and aggregate
- Trained models: Saved models for resumption and feature importance extraction (if model saving enabled)
- Feature importance: Plots and CSV files showing feature importance/coefficients (generated separately)
- Log files: Detailed execution logs (if verbosity > 0)
output_folder/
├── models/ # Trained models (if model saving enabled)
│ └── {Model}_{Target}_models.joblib
├── trained_models/ # Models for cross-dataset validation (from train command)
│ ├── {Model}_{Target}.joblib
│ └── training_metadata.json
├── metrics/ # Detailed performance metrics
│ ├── {target_name}/
│ │ ├── {model_name}_metrics_detailed.csv # Includes Brier Score, ECE, MCE
│ │ └── summary.csv # Summary across all models
│ └── summary_all.csv # Global summary
├── calibration/ # Calibration diagnostics
│ └── reliability_curve_{model}_{target}.png # Reliability curves per fold + aggregate
├── feature_importance/ # Feature importance (if extracted)
│ └── {target_name}/
│ ├── {model_name}_feature_importance.csv # Importance values
│ └── {model_name}_feature_importance.png # Barplot visualization
├── subgroup_analysis/ # Subgroup performance metrics (if configured)
│ └── {target_name}/
│ └── {model_name}_{subgroup_col}_subgroup.csv
├── confusion_matrices/ # Confusion matrix heatmaps
│ └── Confusion_matrix_{model_name}_{target_name}.png
├── report.html # Comprehensive HTML report (includes calibration section)
└── respredai.log # Execution log (if verbosity > 0)
Important
Intended use: ResPredAI is a research software project developed for retrospective data analysis and experimentation with machine learning models for antimicrobial resistance prediction. The software is provided for research and educational purposes only.
See the full history of changes in the CHANGELOG.md file.
If you use ResPredAI in your research, please cite:
@article{Bonazzetti2025,
author = {Bonazzetti, Cecilia and Rocchi, Ettore and Toschi, Alice and Derus, Nicolas Riccardo and Sala, Claudia and Pascale, Renato and Rinaldi, Matteo and Campoli, Caterina and Pasquini, Zeno Adrien Igor and Tazza, Beatrice and Amicucci, Armando and Gatti, Milo and Ambretti, Simone and Viale, Pierluigi and Castellani, Gastone and Giannella, Maddalena},
title = {Artificial Intelligence model to predict resistances in Gram-negative bloodstream infections},
journal = {npj Digital Medicine},
volume = {8},
pages = {319},
year = {2025},
doi = {10.1038/s41746-025-01696-x},
url = {https://doi.org/10.1038/s41746-025-01696-x}
}This research was supported by EU funding within the NextGenerationEU-MUR PNRR Extended Partnership initiative on Emerging Infectious Diseases (Project no. PE00000007, INF-ACT).
Contributions are welcome! Please see CONTRIBUTING.md for guidelines on setting up a development environment, running tests, and submitting pull requests.
This project is licensed under the MIT License - see the LICENSE file for details.