Case Study · AI / Machine Learning
AI Cardiovascular Risk Application
AI-powered cardiovascular risk assessment with clinical inputs, predictive modeling and interpretable health insights.
Offline Python desktop application for estimating heart attack risk from structured patient parameters using a trained neural network.
- ROLE
- Software engineering — machine learning pipeline design, neural-network training, desktop application architecture and local data persistence.
- YEAR
- 2023
TECH_STACK
Technologies
MEDIA
Project media
CONTEXT
Technical context
Estimating heart attack risk from structured patient data requires more than a trained model. The workflow must consistently transform raw inputs, run inference and preserve results in a form that can be reviewed without relying on external services.
FILES
Project files
- View PDF
AI Cardiovascular Risk Application
PDF
PROBLEM
Problem
A trained model alone does not provide a usable risk-assessment workflow. The application needed to prepare structured patient data, apply the same preprocessing used during training, generate a consistent prediction and retain each result for later review within a single offline desktop system.
OVERVIEW
Project overview
This project delivers an offline desktop workflow for estimating heart attack risk from structured patient parameters.
The application guides the user through patient configuration, prediction and review of previously recorded measurements.
A TensorFlow/Keras neural network processes preprocessed cardiovascular data and returns a classification result directly on the local machine.
The model pipeline includes feature preparation, numerical standardization, class balancing with SMOTE and consistent reuse of the trained preprocessing components during inference.
CustomTkinter provides the application interface, while local persistence keeps completed measurements available without requiring an internet connection or external service.
The result is an end-to-end machine learning application that connects dataset preparation, model training, evaluation, deployment and desktop interaction in one system.
APPROACH
Technical approach
- Selected and prepared a structured cardiovascular dataset for supervised classification
- Cleaned and transformed patient features using Pandas and Scikit-learn preprocessing tools
- Standardized numerical inputs and balanced the training data using SMOTE
- Designed and trained a TensorFlow/Keras neural network for heart attack risk prediction
- Applied train/test splitting and early stopping to improve training stability
- Evaluated the model using accuracy, precision, recall, F1-score and a confusion matrix
- Serialized the trained model and preprocessing components for offline application use
- Integrated preprocessing and inference into a CustomTkinter desktop workflow
- Stored completed measurements locally for later review without external services
FOCUS
Technical focus
- TensorFlow/Keras model integration
- Data preprocessing pipelines
- Desktop ML application structure
- Local SQLite persistence
ARCHITECTURE
Architecture notes
- The application separates patient configuration, preprocessing, model inference and result persistence into distinct workflow stages
- Structured patient parameters are transformed using the same preprocessing components that were applied during model training
- Numerical features are standardized before inference to preserve the input distribution expected by the neural network
- The trained TensorFlow/Keras model is loaded locally and executed without external APIs or internet connectivity
- Prediction results are passed from the inference layer to the desktop interface and stored as completed measurements
- The CustomTkinter interface separates navigation, patient setup, risk evaluation and measurement-history views
- Serialized preprocessing objects and model files keep training-time and runtime transformations consistent
- Local persistence allows previously generated measurements and their associated patient parameters to be reviewed later
- The architecture keeps data preparation, inference logic, interface behavior and storage responsibilities separated for easier maintenance
- The complete workflow connects offline model deployment with a user-facing desktop application in a single local system
HIGHLIGHTS
Engineering highlights
- Offline Python application for heart attack risk prediction
- End-to-end workflow from patient parameters to stored prediction results
- TensorFlow/Keras neural network trained on structured cardiovascular data
- Data cleaning, feature preprocessing and numerical standardization
- Training-data balancing with SMOTE
- Model evaluation using accuracy, precision, recall, F1-score and a confusion matrix
- Serialized model and preprocessing components for consistent offline inference
- CustomTkinter interface for patient configuration, prediction and result review
- Local measurement history without external APIs or internet access
- Complete integration of model training, deployment and desktop application workflows
OUTCOME
Outcome
A complete offline desktop application that transforms structured patient parameters into a heart attack risk prediction using a trained neural network. The system combines preprocessing, local inference and persistent measurement history in a single workflow, providing a practical foundation for further model improvement, usability testing and clinical validation.