Sushruta

Inspiration 🧠

Sushruta documents our journey understanding, analyzing, and finally training a model on the UCI AIDS dataset.

We have three goals:

• to make information on healthcare privacy more widely known and accessible 📢⚖️
• to improve AIDS diagnosis outcomes with machine learning 🧠🤖
• to provide interpretability behind the diagnosis process ❤️🖼️

What it does ⚙️

1. SushrutaGPT is a medical & legal chatbot that answers questions about the federal HIPAA healthcare protection act in the United States.
2. Sushruta documents our data visualization of the UCI AIDS dataset.
3. The Sushruta AIDS diagnosis model acheives an accuracy of 83.18% on the UCI dataset. We explain the model's architecture and apply modern interpretability techniques to understand its high performance.

How we built it 💪

Sushruta is built on Streamlit and written in Python. SushrutaGPT uses Google Palm as its language model and embedding engine, ChromaDB as its vector store, and Langchain to coordinate and manage conversation state. The Sushruta AIDS diagnosis model is built using scikit-learn and visualized using seaborn and matplotlib.

Challenges we ran into ⚠️

One major challenge was in improving the performance of the Sushruta AIDS diagnosis model. Initially we could not get higher performance than 60% accuracy. However, after using Gaussian scaling across the feature vectors, we were able to exceed 83%, more than a 20% increase in performance!

Accomplishments that we're proud of 👑

We're very proud of the dramatic increase in performance that we acheived when training our AIDS diagnosis model. This was our first time building a RAG chatbot for a hackathon so we were very proud when all of the different components of SushrutaGPT came together as well.

What we learned 📈

We learned about building reliable, trustworthy chatbots for medical and legal applications, as well as improving the performance of machine learning models iteratively and experimentally.

What's next for Sushruta 🔮

Sushruta's next step will be to increase the performance of the AIDS diagnosis model even further using deeper neural networks and more powerful machine learning models!

Built With

• chromadb
• google-cloud
• google-palm
• langchain
• streamlit