Introduction

Artificial intelligence (AI) is rapidly transforming industries by enhancing decision-making and operational efficiencies. In healthcare, AI has the potential to revolutionize patient care by offering tailored recommendations that help clinicians navigate complex medical cases. In this blog, we’ll dive into the technical and business aspects of building a multi-step AI agent for healthcare, specifically in cardiology.

In collaboration with Dr. Kaustubh Dabhadkar, a practicing cardiologist, we’ll demonstrate how AI can integrate seamlessly into clinical workflows using Azure AI Search, Azure OpenAI, and LlamaIndex. By leveraging these technologies, we aim to create a solution that turns cardiology guidelines from the American College of Cardiology (ACC) into patient-specific decision aids.

Business Context: Why AI in Healthcare?

Healthcare providers often face the challenge of managing vast amounts of clinical data and navigating complex medical guidelines, especially for chronic conditions like atrial fibrillation (AF) and peripheral arterial disease (PAD). AI agents can significantly reduce the time doctors spend researching and cross-referencing guidelines, while also increasing the accuracy and personalization of care plans. This not only improves operational efficiency but also enhances patient outcomes.

Building the Medical AI Agent: Step-By-Step Guide

Step 1: Prerequisites and Setup

Before we dive into the code, let’s get our environment ready. Here are the key technologies you will need to follow along:

• Python 3.9 or higher – The core language used to execute the code.

• Azure AI Search Service – Used to store, index, and retrieve medical guidelines.

• Azure OpenAI Service – Powers the language models for natural language processing.

• LlamaIndex – For document parsing, querying, and structured search.

• CrewAI – Orchestrates multiple agents, each handling different aspects of patient management.

• ReportLab – For generating a patient decision aid in a PDF format.

Let’s start by installing the necessary libraries:

# Core dependencies
! pip install llama-index
! pip install llama-index-core

# Azure-specific dependencies
! pip install azure-search-documents==11.5.1
! pip install llama-index-vector-stores-azureaisearch
! pip install llama-index-embeddings-azure-openai
! pip install llama-index-llms-azure-openai

# Additional tools and utilities
! pip install "crewai[tools]"
! pip install llama-parse
! pip install llama-index-readers-file
! pip install matplotlib
! pip install python-dotenv
! pip install fpdf
! pip install reportlab

These libraries will handle everything from document parsing to querying Azure AI Search and generating PDF reports.

Step 2: Integrating Azure and LlamaIndex

Our AI agent relies heavily on Azure services for vector search and OpenAI for processing natural language queries.We begin by initializing the Azure OpenAI and Azure AI Search services. Next, we use LlamaIndex to parse and index cardiology guidelines from the ACC.

Here’s how to set up Azure OpenAI and LlamaIndex for embedding and searching medical documents:

import os
from dotenv import load_dotenv
from azure.core.credentials import AzureKeyCredential
from azure.search.documents import SearchClient
from azure.search.documents.indexes import SearchIndexClient
from llama_index.core import SimpleDirectoryReader, StorageContext, VectorStoreIndex
from llama_index.embeddings.azure_openai import AzureOpenAIEmbedding
from llama_index.llms.azure_openai import AzureOpenAI
from llama_index.vector_stores.azureaisearch import AzureAISearchVectorStore, IndexManagement
from llama_parse import LlamaParse
import nest_asyncio
nest_asyncio.apply()

load_dotenv()

# Environment Variables
AZURE_OPENAI_ENDPOINT = os.getenv("AZURE_OPENAI_ENDPOINT")
AZURE_OPENAI_API_KEY = os.getenv("AZURE_OPENAI_API_KEY")
AZURE_OPENAI_CHAT_COMPLETION_DEPLOYED_MODEL_NAME = os.getenv("AZURE_OPENAI_CHAT_COMPLETION_DEPLOYED_MODEL_NAME") # I'm using GPT-4o-mini
AZURE_OPENAI_EMBEDDING_DEPLOYED_MODEL_NAME = os.getenv("AZURE_OPENAI_EMBEDDING_DEPLOYED_MODEL_NAME") # I'm using text-embedding-3-small
SEARCH_SERVICE_ENDPOINT = os.getenv("AZURE_SEARCH_SERVICE_ENDPOINT")
SEARCH_SERVICE_API_KEY = os.getenv("AZURE_SEARCH_ADMIN_KEY")
INDEX_NAME = "acc-guidelines-index"
LLAMA_CLOUD_API_KEY=os.getenv("LLAMA_CLOUD_API_KEY")

# Initialize Azure OpenAI and embedding models
llm = AzureOpenAI(
    model=AZURE_OPENAI_CHAT_COMPLETION_DEPLOYED_MODEL_NAME,
    deployment_name=AZURE_OPENAI_CHAT_COMPLETION_DEPLOYED_MODEL_NAME,
    api_key=AZURE_OPENAI_API_KEY,
    azure_endpoint=AZURE_OPENAI_ENDPOINT,
    api_version="2024-06-01"
)

embed_model = AzureOpenAIEmbedding(
    model=AZURE_OPENAI_EMBEDDING_DEPLOYED_MODEL_NAME,
    deployment_name=AZURE_OPENAI_EMBEDDING_DEPLOYED_MODEL_NAME,
    api_key=AZURE_OPENAI_API_KEY,
    azure_endpoint=AZURE_OPENAI_ENDPOINT,
    api_version="2024-06-01"
)

# Initialize search clients
credential = AzureKeyCredential(SEARCH_SERVICE_API_KEY)
index_client = SearchIndexClient(endpoint=SEARCH_SERVICE_ENDPOINT, credential=credential)
search_client = SearchClient(endpoint=SEARCH_SERVICE_ENDPOINT, index_name=INDEX_NAME, credential=credential)

In this step, we set up the Azure OpenAI model and Azure AI Search client, which will be used to query medical guidelines. These guidelines are parsed and indexed using LlamaIndex to streamline the RAG and Agent orchestration.

Step 3: Parsing and Indexing ACC Guidelines

Next, we need to parse cardiology guidelines from the ACC, specifically those related to atrial fibrillation and peripheral arterial disease. This step uses LlamaParse to extract structured data from the PDFs, which we will later index into Azure AI Search.

from llama_parse import LlamaParse

# Set up the parser for ACC PDFs
parser = LlamaParse(result_type="markdown")
documents = SimpleDirectoryReader(
    input_dir="data/acc", 
    file_extractor={".pdf": parser}
).load_data()

# Chunk and Index parsed documents in Azure AI Search
vector_store = AzureAISearchVectorStore(
    search_client=search_client,
    index_name="medical-guidelines-index"
)

index = VectorStoreIndex.from_documents(
    documents, 
    vector_store=vector_store
)

This code parses and loads medical guidelines from PDFs, transforming them into searchable chunks. By using Azure AI Search, we create an efficient and scalable vector index that clinicians can query in real-time.

For businesses, this approach addresses the need to process and retrieve large medical datasets quickly, improving operational efficiency and patient care outcomes.

Dataset Usage

The dataset used in this demo includes ACC guidelines retrieved from the official American College of Cardiology website. These guidelines, focusing on managing atrial fibrillation and peripheral arterial disease, are parsed and integrated into the workflow. This notebook simulates the use case of generating a decision aid for a real cardiology patient under the supervision of a certified medical doctor.

The American College of Cardiology (ACC) is the professional home for the entire cardiovascular care team. ACC's mission is to transform cardiovascular care and improve heart health. The guidelines and information used in this demo are provided by ACC and are subject to their Terms of Service. By using this dataset, you agree to comply with and be bound by the ACC's Terms of Service.

Please note that the use of the ACC's guidelines and information in this demo is for informational purposes only and is not intended as medical advice or as a substitute for the medical advice of a physician. The information contained in this demo may or may not reflect the most current medical developments, and ACC makes no representation with respect to the quality, non-infringement, accuracy, completeness, timeliness, or reliability of such information.

Step 4: Building a Complex Patient Profile

To demonstrate how our AI agent works, we simulate a real-world patient scenario. Sarah Johnson, a 68-year-old woman with a history of atrial fibrillation and peripheral arterial disease, requires personalized treatment recommendations.

Here is the JSON structure for her medical profile:

patient_profile = {
    "patient_id": "43454357890",
    "name": "Sarah Johnson",
    "age": 68,
    "gender": "Female",
    "height_cm": 165,
    "weight_kg": 72,
    "bmi": 26.4,
    "blood_type": "A-",
    "allergies": ["Sulfa drugs"],
    "current_medications": [
        "Flecainide",
        "Apixaban",
        "Metoprolol",
        "Atorvastatin",
        "Cilostazol",
    ],
    "chronic_conditions": {
        "atrial_fibrillation": True,
        "peripheral_arterial_disease": True,
        "hypertension": True,
        "hyperlipidemia": True,
        "coronary_artery_disease": False,
    },
    "family_medical_history": {"heart_disease": True, "stroke": True, "cancer": False},
    "lifestyle_factors": {
        "smoking": "Former smoker, quit 5 years ago",
        "alcohol_use": "Rare",
        "physical_activity_per_week": "4-5 days",
        "diet": "Heart-healthy, low-sodium",
        "sleep_hours_per_night": 6,
    },
    "recent_lab_results": {
        "inr": 1.4, 
        "creatinine_mg_dl": 1.1,
        "ldl_cholesterol_mg_dl": 85,
        "hdl_cholesterol_mg_dl": 55,
        "blood_pressure_mm_hg": "135/80",
    },
    "vaccination_status": {
        "influenza_vaccine": True,
        "covid_vaccine": True,
        "pneumonia_vaccine": True,
    },
    "surgical_history": [
        "Carotid endarterectomy",  
        "Cholecystectomy",         
    ],
    "imaging_history": {
        "last_echocardiogram_date": "2024-07-15",
        "last_carotid_ultrasound_date": "2024-06-20",
    },
    "mental_health": {
        "anxiety": True,
        "depression": False,
        "cognitive_function_issues": False,
    },
    "preferences": {
        "preferred_treatment_type": [
            "Minimally invasive procedures",
            "Evidence-based treatments",
        ],
        "end_of_life_care": "Yes",
        "pain_management": "Non-opioid",
    },
    "recent_visits": [
        {
            "visit_date": "2024-08-25",
            "reason": "Follow-up for atrial fibrillation",
            "notes": "Recurrence of AF despite flecainide. Discussing ablation vs alternative antiarrhythmic options.",
        },
        {
            "visit_date": "2024-07-10",
            "reason": "Peripheral arterial disease management",
            "notes": "Stable claudication symptoms. Continue current management and exercise program.",
        },
    ],
    "af_management": {
        "current_treatment": "Flecainide",
        "treatment_history": [
            "Rate control with metoprolol",
            "Rhythm control with flecainide",
        ],
        "af_recurrence": True,
        "considering_options": ["Catheter ablation", "Alternative antiarrhythmic drug"],
        "chads2_vasc_score": 4,
        "has_bled_score": 2,
    },
    "pad_management": {
        "ankle_brachial_index": 0.78,
        "fontaine_classification": "Stage II",
        "current_treatment": ["Cilostazol", "Supervised exercise program"],
        "last_vascular_assessment": "2024-06-20",
    },
}

We will use this profile to formulate a specific query about Sarah's care:

“Which option reduces stroke risk for a 68-year-old female with atrial fibrillation and peripheral arterial disease who is considering catheter ablation or antiarrhythmic drug therapy?”

Simulating a Real-World Scenario

This profile simulates the comprehensive details a healthcare provider would need to consider in managing a complex patient with cardiovascular disease. From medication management to understanding patient preferences and assessing risk through validated tools, Sarah's profile is representative of real patient cases in cardiology. Healthcare providers can use this information to make informed, evidence-based decisions, such as whether to recommend catheter ablation or adjust anticoagulation therapy based on her risk scores.

Disclaimer

This patient profile is fictional and is used for demonstration purposes only. The data does not represent any real individual, and the use case is designed to illustrate how AI tools can assist in medical decision-making by automating the retrieval of relevant medical guidelines and generating patient-specific recommendations.

Step 5: Query Execution Using Naive RAG

Before moving to a multi-agent system, we’ll use a single-step RAG (Retrieval-Augmented Generation) query engine to generate concise answers directly from the indexed guidelines. This step demonstrates how AI can work like a simple chatbot to provide relevant information.

from llama_index.core.response.notebook_utils import display_response

# Create a query engine
query_engine = index.as_query_engine(llm=embed_model, similarity_top_k=3)

# Perform the query based on Sarah's profile
query_text = (
    f"Which option reduces stroke risk for a {patient_profile['age']}-year-old "
    f"female with atrial fibrillation and peripheral arterial disease, "
    f"who is considering catheter ablation or antiarrhythmic drug therapy?"
)

# Retrieve and display the response
response = query_engine.query(query_text)
display_response(response)

Final Response: For a 68-year-old female with atrial fibrillation and peripheral arterial disease, the recommended option to reduce stroke risk is anticoagulation therapy. Specifically, the CHA2DS2-VASc score should be utilized to assess her risk for adverse outcomes before initiating this therapy. Given her age and the presence of peripheral arterial disease, she would likely qualify for anticoagulation to prevent stroke. Additionally, if she is considering catheter ablation or antiarrhythmic drug therapy, ongoing management with antithrombotic therapy is advised to align with the recommendations for patients with atrial fibrillation.

This code runs a query using the indexed medical guidelines. The Naive RAG approach retrieves a concise answer to the query without diving into deeper multi-agent processes. The response provides a quick and direct answer, giving clinicians an immediate summary of the medical guidelines relevant to Sarah’s condition.

For instance, the response might suggest anticoagulation therapy based on Sarah’s risk scores and medical history, without further analysis.

Synthesis of Naive RAG Response

In Step 5, the Naive Retrieval-Augmented Generation (RAG) query engine worked by leveraging a vector representation of the ACC guidelines previously indexed in Azure AI Search. This allowed us to perform semantic search rather than just relying on keyword matches, ensuring the retrieval of contextually relevant guidelines specific to atrial fibrillation and peripheral arterial disease (PAD).

Here’s what happened under the hood:

• Document Indexing: Prior to running the query, the ACC guidelines and supporting documentation had been converted into vector embeddings using Azure’s OpenAI service. These embeddings capture the semantic meaning of the content, allowing the system to understand the context of medical concepts like “atrial fibrillation” and “anticoagulation therapy” even when expressed in different ways.

• Semantic Search: When the query was executed—asking for management guidelines for a 68-year-old patient with atrial fibrillation and PAD—the system performed a semantic search (specifically, a dense vector search) on the indexed documents. The search engine identified and ranked the most relevant sections of the ACC guidelines based on the embeddings, returning results that not only contained the words used in the query but also aligned with its intent.

• RAG-Driven Response: With these relevant documents in hand, the large language model (LLM) was grounded in authoritative, contextually appropriate medical guidelines. This enabled the model to generate a precise and fact-based response tailored to the patient’s scenario, informing the doctor about the best course of action while ensuring the answer adhered to clinical standards.

Step 6: CrewAI for Multi-Agent Coordination

To automate the retrieval and processing of medical guidelines, we use CrewAI to orchestrate multiple agents working together. Each agent is tasked with a specific role, such as retrieving medical guidelines, translating them into patient-friendly language, developing a treatment plan, and compiling the final output. This multi-agent system allows for efficient and structured AI-driven workflows.

Here’s how the agents are coordinated in this setup:

1. Guideline Expert: This agent retrieves and summarizes relevant ACC guidelines for a specific patient scenario, focusing on treatment options such as catheter ablation or antiarrhythmic therapy.

2. Patient Educator: This agent translates complex medical information into simple, patient-friendly language, ensuring the patient understands the recommendations and treatment options.

3. Treatment Planner: Based on the retrieved guidelines and patient preferences, this agent creates a personalized treatment plan, considering factors like the patient’s medical history and their preference for minimally invasive procedures.

4. Output Generator: This agent compiles the results from the other agents into a structured patient decision aid, formatted into a document that the patient can review.

from crewai import Agent, Task, Crew, Process
from crewai_tools import LlamaIndexTool
from langchain_openai import AzureChatOpenAI

# Define Azure OpenAI configuration for the LLM
azure_llm = AzureChatOpenAI(
    azure_endpoint=AZURE_OPENAI_ENDPOINT,
    api_key=AZURE_OPENAI_API_KEY,
    model=AZURE_OPENAI_CHAT_COMPLETION_DEPLOYED_MODEL_NAME,
    api_version="2024-06-01",
)

# Define the query tool for agents to interact with the ACC guidelines
query_tool = LlamaIndexTool.from_query_engine(
    query_engine,
    name="Medical Guidelines Query Tool",
    description="Use this tool to look up medical guidelines and recommendations.",
)

# We'll use the same patient profile information as before

# Embed patient profile information into the agent goals
# Define the CrewAI agents
guideline_expert = Agent(
    role="Guideline Expert",
    goal=(
        f"Retrieve and summarize relevant ACC guidelines for a {patient_profile['age']}-year-old "
        f"with atrial fibrillation and peripheral arterial disease. "
        "Focus on treatment options such as catheter ablation and antiarrhythmic therapy."
    ),
    backstory="You are an expert on ACC guidelines for managing atrial fibrillation and PAD.",
    tools=[query_tool],
    llm=azure_llm,
    verbose=True,
)

patient_educator = Agent(
    role="Patient Educator",
    goal=(
        f"Translate the medical guidelines into easy-to-understand terms for Sarah Johnson. "
        "Focus on explaining the treatment options available for managing atrial fibrillation."
    ),
    backstory="You explain complex medical terms in patient-friendly language.",
    llm=azure_llm,
    verbose=True,
)

treatment_planner = Agent(
    role="Treatment Planner",
    goal=(
        f"Create a personalized treatment plan for Sarah Johnson, considering her preference for minimally invasive procedures "
        f"and her options of catheter ablation or antiarrhythmic therapy."
    ),
    backstory="You specialize in personalized treatment plans based on patient history and preferences.",
    llm=azure_llm,
    verbose=True,
)

output_generator = Agent(
    role="Output Generator",
    goal=(
        f"Compile the information into a comprehensive patient decision aid document for Sarah Johnson, "
        f"including a clear summary of her treatment options and next steps."
    ),
    backstory="You ensure that the medical recommendations are presented clearly and concisely.",
    llm=azure_llm,
    verbose=True,
)

# Define tasks for each agent
task1 = Task(
    description="Retrieve ACC guidelines for managing atrial fibrillation and PAD.",
    expected_output="Summarized guidelines with a focus on catheter ablation and stroke risk reduction.",
    agent=guideline_expert,
)

task2 = Task(
    description="Translate the guidelines into patient-friendly language.",
    expected_output="Simplified, patient-friendly explanations of the treatment options.",
    agent=patient_educator,
)

task3 = Task(
    description="Personalize the treatment plan according to the patient's preferences.",
    expected_output="A treatment plan tailored to the patient's specific conditions and preferences.",
    agent=treatment_planner,
)

task4 = Task(
    description="Generate a patient decision aid document.",
    expected_output="A final decision aid document summarizing the patient's condition and treatment recommendations.",
    agent=output_generator,
)

# Create the Crew and define the process flow
crew = Crew(
    agents=[guideline_expert, patient_educator, treatment_planner, output_generator],
    tasks=[task1, task2, task3, task4],
    process=Process.sequential,  # Ensures tasks are executed in sequence
    verbose=True,
)

# Execute the multi-step reasoning process
result = crew.kickoff()

# Display the final patient decision aid
print("Final Patient Decision Aid:")
print(result)

Final Patient Decision Aid:
**Patient Decision Aid Document for Sarah Johnson: Managing Atrial Fibrillation**

**Introduction**  
Dear Sarah, this document aims to provide you with clear and comprehensive information about managing your atrial fibrillation (AF). We will discuss your treatment options, necessary lifestyle changes, follow-up care, and important educational resources to empower you in making informed decisions about your health.

---

**1. Treatment Options**  
As someone who prefers minimally invasive procedures, here are your main treatment options for managing atrial fibrillation:

- **Catheter Ablation**  
  - This is a procedure designed to help control or eliminate your AF symptoms. The doctor will use a thin, flexible tube (catheter) to reach your heart and isolate the pulmonary veins. This helps stop the electrical signals that trigger AF. 
  - **Benefits:** Many patients experience significant symptom relief, and it can reduce the frequency of AF episodes, especially if your AF is paroxysmal (comes and goes).
  - **Considerations:** This is a minimally invasive procedure, but it’s important to discuss the potential risks and benefits with your healthcare provider.

- **Antiarrhythmic Medications**  
  - If you decide against or are not a candidate for catheter ablation, you may consider medications such as:
    - **Sotalol:** Helps regulate heart rhythm.
    - **Dronedarone:** Reduces the risk of AF episodes.
    - **Flecainide or Propafenone:** Can be effective in controlling AF but may not be suitable for everyone.
  - **Monitoring:** Regular monitoring is essential while on these medications to ensure they work effectively and safely.

---

**2. Lifestyle Changes**  
To support your treatment and improve your overall heart health, consider these lifestyle adjustments:

- **Heart-Healthy Diet:** Focus on eating plenty of fruits, vegetables, whole grains, and lean proteins. Limit salt and sugar intake.
- **Regular Exercise:** Aim for at least 150 minutes of moderate aerobic activity each week. Consult your doctor for personalized recommendations.
- **Weight Management:** Maintaining a healthy weight can reduce strain on your heart.
- **Stress Management:** Techniques such as meditation, yoga, or deep breathing exercises can help lower stress levels.
- **Sleep Hygiene:** Ensure you get enough quality sleep each night, as poor sleep can affect heart health.

---

**3. Follow-Up Care**  
After your initial treatment, regular follow-up appointments are crucial. During these visits, your healthcare provider will:

- Monitor your heart rhythm to see how well your treatment is working.
- Discuss any side effects from medications or changes in your symptoms.
- Adjust treatment plans as necessary to ensure optimal management of your AF.

---

**4. Patient Education**  
Understanding atrial fibrillation is vital for your health journey. Here are some key points:

- **What is Atrial Fibrillation?**: AF is an irregular and often rapid heart rate that can lead to stroke and other heart-related complications.
- **Importance of Adherence:** Sticking to your treatment plan, whether it’s medication or lifestyle changes, can significantly improve your quality of life.
- **Recognizing Symptoms:** Be aware of symptoms like palpitations, dizziness, or shortness of breath, and report these to your doctor.

---

**Conclusion**  
Sarah, managing atrial fibrillation requires a holistic approach that includes both medical treatment and lifestyle adjustments. This document serves as a guide to help you navigate your options and empower you to engage actively in your health care. Please feel free to reach out with any questions or concerns. Your health is our priority!

--- 

Thank you for taking the time to review this information. Your journey towards managing atrial fibrillation is important, and we are here to support you every step of the way.

How CrewAI Works Under the Hood

In this multi-agent system, CrewAI executes the tasks sequentially, ensuring each agent completes its role before passing the result to the next agent. Here’s what happens step by step:

• Guideline Retrieval: The Guideline Expert agent retrieves ACC guidelines relevant to the patient’s condition, focusing on the management of atrial fibrillation and peripheral arterial disease. The agent refines its query to ensure that the information is both accurate and contextually appropriate for the patient’s case.

• Patient-Friendly Explanation: The Patient Educator agent then takes this information and translates it into easily digestible language. This ensures that the patient can fully understand their treatment options and the implications of each choice.

• Personalized Treatment Plan: The Treatment Planner agent tailors the treatment plan based on the patient’s preferences and medical history, offering a clear decision-making path.

• Final Document Generation: Finally, the Output Generator agent compiles everything into a clear and comprehensive patient decision aid, summarizing the patient’s condition, treatment options, lifestyle recommendations, and follow-up care.

By leveraging CrewAI to orchestrate these agents, the system is able to reduce administrative burden while enhancing clinical decision-making. This not only improves operational efficiency for healthcare providers but also offers personalized, patient-centric care that is both scalable and adaptable.

Step 7: Generating a Patient Decision Aid

Once all the tasks are completed, the AI agent compiles the results into a structured report that Sarah Johnson can take home. Using ReportLab, we automatically generate a patient decision aid in PDF format, summarizing treatment options, lifestyle recommendations, and follow-up care.

The Real-World Impact of AI Agents in Healthcare

By integrating AI agents into healthcare workflows, businesses can enhance operational efficiency and improve patient outcomes. In this scenario, the AI agent reduces the complexity of handling large clinical datasets, providing cardiologists with patient-specific recommendations based on established guidelines.

Future Enhancements

As the medical AI agent evolves, there are several potential enhancements that can further improve its usability and integration into clinical workflows. Below are a few ideas for future improvements:

• Real-Time Chat Interface:

  • Develop a chat interface for direct communication with patients. This could be integrated into existing healthcare portals or even through WhatsApp to offer patients quick, accessible guidance and follow-up.

• Automated Email Notifications:

  • Set up an automated system that emails patients their decision aid after a consultation, ensuring they have a digital copy of the recommendations, along with additional educational materials.

• Patient Feedback Loop:

  • Implement a feedback mechanism where patients can upvote/downvote the recommendations they received or provide free-form feedback. This feedback can be integrated into a learning loop, continuously improving the AI system.

• Expanding ACC Guidelines with Real-World Data:

  • Use patient feedback and new case studies to expand the ACC guidelines stored in the system. This would allow the AI agent to reference more real-world examples and improve the quality of insights provided to doctors for new patients with similar profiles.

These enhancements can help scale the AI agent into a comprehensive healthcare solution, offering real-time, personalized care that improves patient engagement and empowers clinicians with actionable insights.

Conclusion

This project showcases how AI agents can transform healthcare by integrating data retrieval, multi-agent coordination, and natural language processing to assist doctors in clinical decision-making. Developed in collaboration with Dr. Kaustubh Dabhadkar, this AI agent automates the retrieval of cardiology guidelines and personalizes care recommendations, delivering clear value in clinical settings.

Healthcare businesses that invest in AI-driven systems like this one can expect significant improvements in care delivery, operational efficiency, and patient satisfaction.

References

• azure-ai-search-python-playground/azure-ai-search-crewai.ipynb at main · farzad528/azure-ai-search-python-playground (github.com)

• crewAI - Platform for Multi AI Agents Systems

• LlamaIndex - LlamaIndex

• Introduction to Azure AI Search - Azure AI Search | Microsoft Learn

• Home - American College of Cardiology (acc.org)

If you’re interested in exploring AI solutions tailored to your business needs, feel free to reach out. Let’s collaborate to build AI-driven tools that add real-world value.