Bass: AI-SDA Generator

Gordan Milovac, • JavaScriptReact.jsHTMLCSSInterSystems ObjectScriptInterSystems IRIS
Back

This project is a final project for InterSystems Sales Engineering Summer 2024 Internship. I was mostly working on the AI Prompt Engineering, UI Design and Documentation.

Main Page

Project details

Creating and populating patient data in the Summary Data Architecture (SDA) format is a complex and time-consuming task, often taking over a week to complete manually. SDA, developed by InterSystems, is a comprehensive data model used within the Unified Care Record (UCR) to standardize and manage healthcare data, ensuring seamless integration and interoperability across different systems.

The AI-SDA Generator project significantly reduces the time needed to create and populate SDA patient records by leveraging AI technology. Using a user-friendly chat interface, users can interact with and edit the SDA data model using natural language. This project aims to provide efficient data interaction, accurate data transformation, seamless data integration, enhanced user experience, and secure data handling.

Github repo:

https://github.com/Wzelong/AI-SDA

Project Presenatation:

https://youtu.be/1xkQfhAn08Q

All About That Bass (ad):

https://youtu.be/owz7UZDT2ms

Create

Project Description

This project is a final project for the Sales Engineering Summer 2024 Internship at InterSystems. It utilizes AI technology to generate and populate demo SDA medical records. The User Interface is built using React framework and the database is Google Firebase. Data population is handled using AI (ChatGPT) and careful prompt engineering. The file storage and transformation is handled using InterSystems HealthShare. Users get to specify certain constraints (case, specification, encounters...) and use them to create demo patient medical records. This can be done manually or using the autogenerate feature which sends parallel API requests to the AI. The medical records are stored in the History tab, and can be edited or deleted.

Timeline

Design Choices

User Interface:

The design of the user interface incorporates InterSystems' brand colors, ensuring a professional and cohesive look:

export const theme = {
  colors: {
    darkPrimary: '#333694',
    lightPrimary: '#EAEBFA',
    darkSecondary: '#00A39E',
    lightSecondary: '#BEE4E3',
    darkText: 'white',
    lightText: '#333694'
  }
}

The interface is designed for simplicity and usability, making it easy for users to fill out mandatory data fields efficiently. The chat interface allows for natural language interaction, making the system intuitive and user-friendly.

Forms:

We use Ant Design to create intuitive and user-friendly forms. AntDesign’s robust form components ensure all necessary information is easily accessible and editable. This contributes to a smooth user experience, allowing users to quickly input and modify patient data.

Tables:

AG Grid is employed to create dynamic and responsive tables capable of handling large datasets efficiently. AG Grid's features, such as sorting, filtering, and grouping, provide users with a clear and organized view of patient data. This enhances data management and accessibility, allowing users to efficiently review and edit large amounts of information.

Parallel API Calls:

To optimize performance and efficiency, the system uses batched AI calls for parallel processing. This approach combines outputs from multiple API calls, significantly reducing the time required to generate complete patient records. The parallel processing capability ensures that the system can handle multiple requests simultaneously without compromising speed or accuracy.

Prompt Engineering:

Careful prompt engineering is a critical component of this project. Prompts are meticulously designed to ensure that the AI generates relevant and accurate data that fits seamlessly into the SDA model. This ensures that the generated data is of high quality and meets the specific requirements of the healthcare domain.

Population:

We decided to generate patient data using predefined medical cases and appending them randomly to a base case. This method allows us to create up to 10,000 unique patient records in under 10 seconds. Users can choose the level of diversification of the results compared to the base patient, ensuring a tailored and relevant dataset. The results are then displayed in pie charts for easier visualization, providing a clear overview of the generated patient data. This approach not only ensures quick data generation but also offers flexibility in data variety, enhancing the usability and applicability of the patient records.

Inspect

Team members and contributions

Zelong Wang (wzelong): React Functionality, UI Design, Parallel API Calls

Gordan Milovac (gmilovac): AI Prompt Engineering, UI Design, Documentation

Time allocated: 400+ hours

Errors / bugs

This code is very AI dependent. Using careful prompt engineering and file validation, we made sure to minimize the margin of error... but it still exists. AI will sometimes "hallucinate". Generated data is meant to be inspected before the final submission.

Setup

Build and run your program:

You can now use our React app!

Use different AI model:

You can now use a different AI model!

Population

Tutorial and Functionality

YouTube Tutorial: https://youtu.be/LDagH8bnXjE

1. Creating a Patient:

You now have your own Patient!

2. Creating a Timeline:

You now have your own Timeline!

3. Creating an SDA Table:

You now have your own SDA Table!

4. Creating an XML SDA File:

You now have your own XML SDA file!

5. (Optional) Creating Population:

You now have your diversified Population!

6. (Optional) Redo:

You just generated another Patient/Population!

Help

Extra

Chatbot:

We built a bot that can assist you in the patient creation process!

You can now use the InterSystems chatbot!

Tips with the Chatbot:

  1. Clicking on the paperclip icon allows you to import an already existing XML SDA file

  2. If you say "Take me to" and then specify the form, it will take you to that form

  3. If you say "Imagine a patient" or "Create a case" and give further instructions, it will create a medical case for you and auto-populate the 'Design/Requirements' Form

  4. If you say "How to" and then specify the step you are struggling with or don't understand, it will search the documentations and give you an answer

  5. If you ask any medical or technical question if will use the selected AI model to answer

  6. (EE) If you ask "Who created this app" it will respond and give you a quick info about the creators

History: Your generated patients will be stored in the history tab. View them, edit them or delete them as you wish!

You can now view and update old SDA files!

© Gordan Milovac.Resume PDF