Sylva AI

• 

  export patient trend tab

• 

  credits tab

• 

  chatbot tab

• 

  data visualization tab

Inspiration

Our inspiration for Sylva AI came from seeing the impact that cardiovascular diseases and heart attacks have around the world. As highschool students we were mainly motivated by the idea that technology like AI could potentially help predict events before they even happen and since heart disease is responsible for nearly one/third of all global death and so many of its early warnings are missed we wanted lean into how ML can turn basic medical data to something that can turn into a early warning system.

What it does

It's an interactive web-based design that predicts how likely a heart attack is based on the diagnostic input data. Users can enter details like age, cholesterol, blood pressure, chest pain type, fasting blood sugar, ECG results etc. and this app analyzes this data and outputs a probability score of heart attack risk which is divided into 3 levels which are: Low, Moderate and High and this also has a patient risk tracker over time and even a detailed breakdown of which features matter most in the prediction process and this runs in real time.

How we built it

We built Sylva AI using Python as our core programming language and Streamlit for creating the web interface, and for the machine learning part, we used the scikit learn library to train a LGBM Classifier, which performed best in terms of accuracy and AUC score. To address class imbalance in the dataset (since heart attacks are less frequent than non-heart attack cases), we used SMOTE (Synthetic Minority Oversampling Technique), which allowed us to train a better model. Our dataset was a combination of open-source heart disease datasets like UCI Heart Disease and Cleveland Clinic datasets, and we also added synthetic data values for HDL, LDL, and BMI to enhance its prediction accuracy of the model and the inputs were normalized using StandardScaler. We also experimented with a neural network using TensorFlow and Keras by building a 5-layer deep learning model, but for final predictions, we decided to choose the LGBM model due to its consistency and faster training time The chatbot was built with groq's llm API"" **The chatbot was built using the help of chatgpt and synthetic data was initially created by us we used chatgpt to add more features

Challenges we ran into

A challenge was balancing the dataset, and without SMOTE, the model tended to favor more non-heart attack predictions, which reduced its real-world usefulness. Another challenge was that we had to carefully manage dependencies and figure out how to load both types of models efficiently, and designing a professional-looking and user-friendly interface was also more complicated than we expected, especially while using Streamlit, which has layout limitations. There was also a major challenge of communication since we were both in different countries/timezones.

Accomplishments that we're proud of

We are proud that we were able to develop a fully functioning/working AI powered heart attack risk predictor, especially as high school students, and despite our limited experience, we managed to create a clean, responsive, and a medically informed interface. We successfully built a machine learning pipeline from raw data to final prediction, which incorporates our model evaluation metrics like AUC, ROC curves, and confusion matrices

What we learned

We learned how machine learning can be applied to real-world health challenges and how important data preprocessing is for accurate predictions, and we also gained experience with feature scaling, data balancing, and model evaluation. We also explored the differences between traditional machine learning algorithms and deep learning models, and how each has strengths and weaknesses depending on the problem.

Built With

• csv
• groq
• joblib
• keras
• matplotlib
• numpy
• pandas
• pickle
• plotly
• python
• scikit-learn
• seaborn
• smote
• streamlit
• tensorflow