Important Topics Studied

  • Digital Healthcare
    • FHIR (Fast Healthcare Interoperability Resources)
    • OMOP (Observational Medical Outcomes Partnership)
    • Clinical Oncology
    • Generative AI in Healthcare
    • Natural Language Processing (NLP) in Healthcare
    • Data Interoperability and Standards
    • Electronic Health Records (EHR)

Techstack

  • Python
  • Streamlit
  • R
  • Docker
  • PostgreSQL
  • HAPI FHIR
  • LLM
    • OpenAI - gpt-4o-mini
    • GoogleAI - gemini-2.0-flash
  • Langchain/Langgraph

Projects Developed

Throughout this course we have developed many mini projects which resulted in a large project around Care Compliance Dashboard. We also had Final Group Project where we developed the project of “MIPS (Merit Based Incentive Payment System) Quality Measure Calculation using OpenAI”

Here is the labs details in this course:

Labs: https://leap-of-faith-technologies.gitbook.io/cs-595-digital-healthcare-informatics-and-ai

GitHub Link (Opensource!): https://github.com/haard7/OpenAI-MIPS

Project-1: AI Powered Care Compliance Dashboard - (Consisting of multiple labs)

Lab-1: Setup Care Compliance Dashboard

  • It contained the setting up of the whole Care Compliance Dashboard developed by LOF (Leap of Faith) Technologies. It is the centerpiece of the whole digital healthcare stack. We used Docker to setup the whole project. It contained 4 different containers representing 4 different microservices including HAPI FHIR server and PostgreSQL database.

Lab-2: FHIR Integration to CCD

  • After getting the theoretical idea about FHIR from Dr. Bob Dolin , we started practically understanding how exactly creating the patient record interact with HAPI FHIR Server. We also performed the CRUD operation using Postman to interact with core FHIR resources.

Lab-3: Clinical Ontology Tokenization

  • In this lab I worked on clinical ontologies (e.g., SNOMED CT, ICD-10, LOINC, CPT) and their role in ensuring semantic interoperability in healthcare systems. We used LLMs (Large Language Models) to process clinical note, extract clinical concepts, map them to the appropriate ontology codes, and identified the medications. We used two approaches. One using IMO Tokenization and another using LLM gemini-2.0-flash. Finally we compared the results on streamlit UI.

LAB-4: Cohort Discovery and Analysis

  • Completed a capstone lab project focused on transforming and integrating healthcare data from multiple sources–including EHRs, HIEs, FHIR resources, tokenized clinical notes, and patient engagement data–into the OMOP Common Data Model (CDM). Gained hands-on experience with the structure and purpose of the OMOP CDM, executed ETL (Extract, Transform, Load) processes to standardize the data, and utilized the standardized data for analytics and cohort analysis. The project highlighted the critical role of data standardization in enabling research, interoperability, and large-scale healthcare analytics.

Lab-5: Multi Agent FHIR assistant

  • This was very interesting lab where we had to develop the multiple agents with different tool calling in order to talk to FHIR server effectively and respond with summary on streamlit UI. We used OpenAI Agent SDK for developing the agent in Python

Miscellaneous:

  • We also worked on other labs which included working with different healthcare technologies and AI tools. For example, healthcare gorilla for better interoperatibility, D-ID for AI based interactive avatar used in chatbot to improve the custoer experience and engagement.

Project-2: MIPS (Merit Based Incentive Payment System) Quality Measure Calculation using OpenAI

About:

This was our group project where we had to develop the MIPS Quality Measure Calculation using OpenAI. We used Langchain and Streamlit to develop the project. We used gpt-4o-mini in our project.

Project Demo: Click Here

Techstack:

▶ Langchain - Framework for building applications with LLMs
▶Streamlit - Framework for building interactive web applications
▶ OpenAI - Used its api to use gpt-4o-mini
▶Pinecone - Vector database for storing and retrieving embeddings and patient info.

Project Overview:

  • Every year the Centers for Medicare & Medicaid Services (CMS) releases a set of quality measures that healthcare providers must report on to demonstrate their commitment to providing high-quality care. These measures are part of the MIPS program, which aims to improve patient outcomes and reduce healthcare costs.
  • Clinics and Hospitals get Incentives for medicare based on the performance. So better the MIPS quality measure reporting, better the incentives.
  • So we came up with the idea that we can accurately calculate the MIPS quality mesures using openAI model, it would reduce the Manual Validation and Reduce the time to prepare and report the MIPS quality measures.

Development Flow

  • As we had only couple of weeks to work on this exciting project, we started with single quality measure (ACEP22) which is related to pulmonary embolism. For single quality mesure we used the method of Prompt Injection to develop the prototype application. It was giving correct answer by analyzing the requirements for “Demominator” and “Numerator” and then we calculated final Percentage of peformance for that quality measure by analyzing the patient note using LLM (Large Language Model)

  • Then, we gradually added another quality measure ACEP60. It worked well with the same prompt injection method. But we were not able to use the true power of vectorDB and RAG (Retreival Augmented Generation) method. So we decided to store the quality measure and leverage the power of RAG. We presented our demo with “Prompt Injection” method and presented in form of Value Proposition along with Demo.

Future work:

  • Leveraging the RAG method to automate and expand the calculation and analysis of MIPS quality measures would be the next step. Our course was finished but we are optimistic to expand this code and implement the RAG pipeline in given scenario.

Thank You….