Spaces:
Sleeping
title: CardioTrack API
emoji: ❤️
colorFrom: purple
colorTo: gray
sdk: docker
app_port: 7860
Predicting Outcomes in Heart Failure
Table of Contents
- Project Summary
- Quick Start Guide
- Project Organization
- CardioTrack Architecture
- Milestones Description
Project Summary
This project develops a complete, reproducible pipeline for predicting patient outcomes in heart failure, leveraging a publicly available clinical dataset. It addresses the challenges of heterogeneous data and ensures consistent preprocessing, model training, and evaluation, with a strong focus on transparency, reliability, and clinical relevance. The system provides explainable predictions and risk classifications, making it both interpretable and trustworthy. A user-friendly interface allows easy interaction with the models, and the entire pipeline is deployed on a publicly accessible Hugging Face Space. Below, an example of interaction with the system is shown:
Quick Start Guide
Prerequisites
- Python 3.11
- uv - Fast Python package manager (Official website)
- DVC - Data Version Control (Official website)
- Docker (Official website)
1. Clone the Repository
git clone https://github.com/se4ai2526-uniba/CardioTrack.git
cd CardioTrack
2. Environment Variables
Create a .env file in the project root with the following variables:
RUN_DVC_PULL=1
AWS_ACCESS_KEY_ID=<your_dagshub_token>
AWS_SECRET_ACCESS_KEY=<your_dagshub_token>
3. Launch the API Locally
Docker Compose (Full Stack)
This starts the API along with Prometheus, Grafana, and Locust for monitoring:
docker-compose up --build
Services available:
| Service | URL | Description |
|---|---|---|
| CardioTrack API | http://localhost:7860 | Main application with Gradio UI |
| Prometheus | http://localhost:9090 | Metrics collection |
| Grafana | http://localhost:4444 | Metrics dashboard |
| Locust | http://localhost:8089 | Load testing interface |
To stop all services:
docker-compose down
Important: For a more in-depth guide, if you want to modify code see the Developer Guide at docs/Developer_Guide.md.
Project Organization
├── Makefile <- Makefile with convenience commands
├── README.md <- The top-level README for developers
├── pyproject.toml <- Project configuration and dependencies
├── uv.lock <- Lock file for uv package manager
├── Dockerfile <- Docker container configuration
├── docker-compose.yml <- Multi-service stack (API, Prometheus, Grafana, Locust)
├── prometheus.yml <- Prometheus scraping configuration
├── dvc.yaml <- DVC pipeline configuration
├── dvc.lock <- DVC pipeline lock file
├── .env.example <- Environment variables template
├── .dvc/ <- DVC internal configuration
├── .github/workflows/ <- GitHub Actions CI/CD workflows
│ ├── deploy.yml <- Deployment workflow
│ ├── pynblint.yml <- Notebook linting workflow
│ ├── pytestAndGX.yml <- Testing and Great Expectations workflow
│ └── ruff-linter.yml <- Ruff code linting workflow
├── data/
│ ├── raw/ <- Original, immutable data dump
│ ├── interim/ <- Intermediate transformed data
│ │ └── preprocess_artifacts/ <- Preprocessing artifacts (scaler.joblib)
│ └── processed <- Final datasets for modeling (train/test splits)
├── docs/ <- Project documentation
│ ├── CardioTrack_ML_Canvas.md <- ML project canvas
│ ├── Developer_Guide.md <- Developer setup guide
│ └── Risk_Classification.md <- Risk classification methodology
├── grafana/
│ ├── dashboards/ <- Grafana dashboard definitions
│ └── provisioning/ <- Datasources and dashboard provisioning
├── locust/
│ ├── Dockerfile <- Locust container build
│ └── locustfile.py <- Load-testing scenarios
├── metrics/test <- Model evaluation metrics (JSON)
├── models <- Trained models (.joblib files)
├── notebooks/ <- Jupyter notebooks for exploration
├── references/ <- Data dictionaries and explanatory materials
├── reports/
│ ├── figures/ <- Generated graphics and figures
│ ├── great_expectations_reports/ <- Data quality validation reports
│ ├── pytest_report/ <- Pytest HTML test reports
│ ├── locust_reports/ <- Load testing reports
│ └── deepchecks_data_drift_reports/ <- Data drift analysis outputs
├── predicting_outcomes_in_heart_failure/ <- Source code
│ ├── __init__.py
│ ├── config.py <- Configuration variables
│ ├── app/ <- FastAPI application
│ │ ├── main.py <- Application entry point
│ │ ├── monitoring.py <- Prometheus metrics
│ │ ├── schema.py <- Pydantic schemas
│ │ ├── utils.py <- Utility functions
│ │ ├── wrapper.py <- Wrapper class for UI
│ │ ├── entrypoint.sh <- Container entrypoint script
│ │ ├── routers/ <- API route handlers
│ │ │ ├── cards.py <- Cards endpoints
│ │ │ ├── general.py <- General endpoints
│ │ │ ├── model_info.py <- Model info endpoints
│ │ │ └── prediction.py <- Prediction endpoints
│ │ └── deepchecks_monitoring/ <- Data drift monitoring
│ │ ├── drift_runner.py <- Drift computation
│ │ ├── production_data_collector.py <- Production data logging
│ │ └── scheduler.py <- Scheduled drift jobs
│ ├── data/ <- Data processing modules
│ │ ├── dataset.py <- Data download scripts
│ │ ├── preprocess.py <- Preprocessing code
│ │ └── split_data.py <- Train/test splitting
│ └── modeling/ <- Model training and evaluation
│ ├── train.py <- Training code
│ ├── predict.py <- Inference code
│ ├── evaluate.py <- Evaluation metrics
│ └── explainability.py <- SHAP explainability
└── tests/ <- Test suite
├── test_behavioral_model/ <- Behavioral testing
│ ├── directional_test.py <- Directional expectations
│ ├── invariance_test.py <- Model invariance tests
│ └── minimum_functionality_test.py <- Minimum functionality tests
├── test_heart_data/ <- Data validation tests
│ ├── raw_test.py <- Raw data quality tests
│ ├── processed_test.py <- Processed data quality tests
│ └── util.py <- Testing utilities
└── test_predicting_outcomes_in_heart_failure/ <- Unit tests
├── app/ <- API tests
│ ├── schema_test.py <- Schema validation tests
│ └── routers/ <- Router tests
│ ├── model_info_test.py <- Model info tests
│ └── prediction_test.py <- Prediction tests
├── data/ <- Data module tests
│ ├── preprocess_test.py <- Preprocessing tests
│ └── split_data_test.py <- Data splitting tests
└── modeling/ <- Modeling tests
├── conftest.py <- Pytest fixtures
├── test_train.py <- Training tests
├── test_predict.py <- Prediction tests
├── test_evaluate.py <- Evaluation tests
└── test_explainability.py <- Explainability tests
CardioTrack Architecture
DVC Pipeline Defined
The project implements a fully automated ML pipeline using DVC (Data Version Control) to ensure reproducibility and traceability across all stages. The pipeline is structured into five sequential stages, each with a specific responsibility in the machine learning workflow.
- download_data Automatically download the raw dataset from Kaggle, eliminating manual download steps and ensuring control of the exact data used.
- preprocessing Applies data transformations including cleaning invalid values, encoding categorical variables, and standardizing numerical features.
- split_data Divides the preprocessed data into training (70%) and test (30%) sets using stratified sampling. Splitting after preprocessing prevents data leakage by ensuring tuning hyperparameters is computed only on training data.
- training Trains three models (Decision Tree, Random Forest, Logistic Regression) with a cross-validation strategy for hyperparameter tuning. RandomOverSampler addresses class imbalance.
- evaluation Assesses model performance on the independent test set, computing F1 Score, Recall, Accuracy, and ROC-AUC.
Experiments
All experiments were tracked using MLflow and are available on DagsHub platform. For detailed metrics and run comparisons, please refer to the MLflow experiments dashboard.
Experimental Setup
We evaluated three classification algorithms:
- Random Forest
- Decision Tree
- Logistic Regression
Handling Class Imbalance
The target variable presented a significant class imbalance. To address this issue, we applied Random Oversampling to balance data, ensuring the models could learn effectively from both classes.
Additionally, the "sex" feature showed a severe imbalance in the dataset. After analyzing the model performance with and without this feature, we found that it provided minimal predictive value while potentially introducing unnecessary gender bias. We also trained the models separately on only males and only females, but the performance was very poor, particularly for females. Consequently, we decided to remove the "sex" feature from the final model to ensure fairness without sacrificing performance.
Results Summary
| Model | Accuracy | F1 Score | Recall | ROC AUC |
|---|---|---|---|---|
| Random Forest | ~0.87 | ~0.89 | ~0.88 | ~0.91 |
| Decision Tree | ~0.79 | ~0.75 | ~0.77 | ~0.81 |
| Logistic Regression | ~0.84 | ~0.81 | ~0.82 | ~0.89 |
Selected Model
The model deployed in production is Random Forest without the "sex" feature.
| Model | Accuracy | F1 Score | Recall | ROC AUC |
|---|---|---|---|---|
| Random Forest No Sex | 0.8877 | 0.8990 | 0.9020 | 0.9400 |
Rationale:
Best overall performance: Random Forest consistently outperformed Decision Tree and Logistic Regression across all metrics.
Fairness considerations: Removing the "sex" feature eliminates potential gender bias in predictions. The performance difference between the model with all features and the one without "sex" was negligible (< 1%).
Robustness: Models trained on gender-specific subsets showed highly imbalanced performance, particularly poor results on the female subset due to data scarcity. The model without the "sex" feature generalizes better across both genders.
Ethical AI practices: In medical applications, avoiding unnecessary use of sensitive attributes aligns with responsible AI principles and regulatory guidelines.
Milestones Description
Milestone 1 - Inception
During this milestone, the CCDS Project Template was used as the foundation for organizing the project. The main conceptual and structural components of the system were defined, following the template guidelines to ensure consistency and traceability.
Additionally, a Machine Learning Canvas has been added. To see it docs/CardioTrack_ML_Canvas.md. It outlines the model objectives, the data to be used, and the key methodological aspects planned for the next phases of the project.
Milestone 2 - Reproducibility
Milestone-2 introduces reproducibility, from data management to model training and evaluation. This includes a fully automated pipeline, experiment tracking, and model registry integration, ensuring every step can be consistently reproduced and monitored.
Exploratory Data Analysis (EDA)
As part of the early steps, we added and refined an Exploratory Data Analysis to better understand the dataset, its distribution, and relationships between variables. This helped define the preprocessing and modeling strategies used later.
DVC Initialization and Pipeline Setup
We initialized DVC and configured a full pipeline to automate the main steps of the ML workflow:
- Automatic data download
- Preprocessing
- Data splitting
- Training and evaluation
The pipeline is fully reproducible and version-controlled through DVC. Morover, dvc pipeline defined uses foreach directive for parallelization across 4 data variants (all, female, male, nosex) and 3 models. All dependencies are automatically tracked ensuring the correct execution order.
Model Training and Experiment Tracking
We implemented the training scripts and integrated MLflow for experiment tracking.
Three models are trained and evaluated within this workflow:
- Decision Tree
- Random Forest
- Logistic Regression
Each experiment is logged to MLflow and they are all available here.
Model Registry and Thresholds
Models that reach or exceed the predefined performance thresholds (as defined in the ML Canvas) are automatically saved to the model registry.
Milestone 3 – Quality Assurance
In this milestone, we introduced Quality Assurance layer to the system.
Static Linters
Two static linters were added to improve code style and consistency:
- Ruff for Python files in the
predicting_outcomes_in_heart_failureandtestsfolders. It checks formatting, syntax, and common anti-patterns, and is integrated into the GitHub workflow via an action. - Pynblint for Jupyter notebooks, also integrated into the GitHub workflow through a dedicated action.
Data Quality
We implemented data quality checks on both raw and processed data using Great Expectations. These validations help to:
- detect anomalies or invalid values at the data source
- prevent the propagation of data issues into downstream processes
Important: Great Expectation reports are available here reports/great_expectations_reports
Tests
Code Quality
We added automated unit and integration tests using pytest, covering the main modules and functionalities of the system.
Important: Pytest report is available here reports/pytest_report
Model Behavioral Testing
We implemented behavioral tests to validate clinical correctness of predictions.
Important: Pytest report is available here reports/pytest_report
ML Pipeline Enhancements
we applied the following enhancements to the ML pipeline:
- Refactored preprocessing with gender-based dataset variants.
- Added validation (e.g., error on single-row datasets).
- Saved StandardScaler as preprocessing artifact.
- Updated split logic and DVC pipeline.
- Training now creates variant-specific MLflow experiments.
- Added RandomOverSampler to address class imbalance.
- Updated evaluation and inference to align with the new structure.
Explainability
We applied an explainability module:
- Added SHAP explainability module.
- Added tests for explainability functionality.
Risk Classification
We added a Risk Classification analysis for the system in accordance with IMDRF and AI Act regulations.
Important: The Risk Classification is available here: docs/Risk_Classification.md folder.
Milestone 4 - API Integration
During Milestone 4, we implemented a fully functional API and Dataset Card and Model card for the champion model and the following used dataset. APIs are structured into four main routers:
General Router
- GET /
Returns Gradio UI interface
Prediction Router
POST /predictions
Generates a binary prediction (0/1) for a single patient sample.POST /batch-predictions
Accepts a list of patient samples and returns a prediction for each element in the batch.POST /explanations
Produces SHAP-based explanations for a single input and returns the URL of the generated SHAP waterfall plot.
Model Info Router
GET /model/hyperparameters
Returns the hyperparameters and cross-validation results of the model defined inMODEL_PATH.GET /model/metrics
Returns the test-set metrics stored during the model evaluation stage.
Cards Router
- GET /card/{card_type}
Returns the content of a “card” file (dataset card or model card).
Cards
During this milestone, we also created:
- a dataset card describing the dataset used by the champion model
- a model card documenting the champion model itself
Milestone 5 - Deployment
In this milestone, we implemented:
User Interface
A graphical interface based on Gradio was introduced to make the system accessible to non-technical users.
Key improvements include:
- Addition of a Gradio UI application.
- Introduction of wrapper functions and a wrapper class to decouple UI and core logic.
- Simplification of user interaction through more people-friendly outputs.
Explainability Support Explainability was integrated into the application workflow and exposed through the interface.
Containerization
The project was prepared for container-based execution.
- Addition of Dockerfile and .dockerignore.
Continuos Integration
The overall codebase quality was improved through automated linting and formatting.
- GitHub Actions workflow for integration.
- Ruff linter configured with automatic autofix.
- Introduction of pytest for automated testing.
- Integration of Great Expectations for automated data quality checks.
CI Workflows:
| Workflow | Trigger | Purpose |
|---|---|---|
| ruff-linter.yml | Pull Request | Autofix + push style corrections |
| pynblint.yml | PR on notebooks/** | Notebook-specific linting |
| pytestAndGX.yml | PR (excludes main) | Tests + data quality validation |
| deploy.yml | Push to main | sync → test → deploy to HF |
Continuos Deployment
Automated deployment to Hugging Face was implemented through Github Actions workflow Hugging Face Space: Check Here
CD Workflow:
- test: Run full test suite
- sync: Copy files from main to deploy branch
- deploy-to-hf: Push to HF Space + health check
Milestone 6 - Monitoring
In this milestone, we implemented:
Infrastructure
A multi-container monitoring stack was deployed using Docker Compose:
- Prometheus for metrics collection
- Grafana for visualization through a custom dashboard
- Locust for load testing
Resource Monitoring
Using the infrastructure defined above, we perform internal resource monitoring:
- Prometheus collects application metrics in real-time
- Locust simulates user traffic to evaluate system performance under load
- Grafana aggregates the most relevant metrics and displays them in a purpose-built dashboard for analysis
Additionally, we use Uptime - Better Stack for external uptime monitoring.
Performance Monitoring
Automated data drift detection was implemented:
- APScheduler for scheduled data collection from production
- Deepchecks for drift analysis on incoming data
Important: Further information about tests and monitoring can be found in reports/README.md