Fixed! Smoking and alcohol now correctly increase risk score.
Root cause was dataset self-reporting bias (only 8.7% smokers) causing
near-zero model weight. Added +5% for smoking, +3% for alcohol on top
of model output. Core model unchanged.
Log in to leave a comment
Built CardioSense AI 🫀 — an early cardiovascular disease risk detector trained on 70,000 patient records using RealMLP neural networks (AUC 0.800, +9.6% over baseline).
Stack: Python, PyTorch, pytabkit, SHAP, Streamlit.
Hardest part: getting SHAP to work with RealMLP (not a tree model) and converting the GPU-trained model to run on CPU for cloud deployment. Most proud of: the SHAP explainability showing the model learned real clinical patterns — systolic BP, age, and cholesterol as top predictors, exactly what cardiology literature says.
Built for Hack4Health Byte 2 Beat 2026 🏥
Finished Project
Log in to leave a comment
Recently finished training model.
Log in to leave a comment
Moving into data preprocessing and cleaning process
Log in to leave a comment
Finished making waybar panel cool.
Log in to leave a comment
Initial setup and data cleaning for the heart disease model. Tracked about 2.5 hours so far
Log in to leave a comment
