---

The code for the growth chart app is forked from SMART on FHIR and extends it with an LLM.

Abstract

Monitoring child’s growth is essential for identifying and treating disorders and diseases early that manifest through abnormal growth patterns. We explore the integration of an Large Language Model (LLM) into a simulated clinical setting to assist primary care pediatricians in detecting growth abnormalities. The focus is on extending the open-source SMART on FHIR Growth Chart App by developing an AI tab that provides an analysis of patient growth data and medical history, helping primary care pediatricians make informed decisions about specialist referrals.

Using a User-Centered Design (UCD) approach, this study conducted semi-structured interviews with pediatricians to gather requirements and iteratively refine the AI tab. Synthetic patient data, created through a literature search of common growth disorders and diseases, was utilized to be able to use the app and to generate responses with the LLM. The AI tab was implemented using the OpenAI Assistans API with Retrieval Augmented Generation (RAG) and tested for usability and functionality with three pediatricians in an evaluation. In addition, in this evaluation the responses of the model for the growth abnormality detection, patient history analysis, recommended specialist referral, differential diagnosis and executive summary were assessed for both relevance and correctness from the pediatrician's.

The results illustrate how an LLM can be technically integrated into a simulated clinical setting. The evaluation highlighted that a high level of usability was achieved through the UCD approach, and the content selected for the AI tab is tailored to the needs of the users. The model correctly analyzed the growth of three out of five synthetic patients. However, the other responses were insufficient, particularly the recommended specialist referrals and differential diagnoses.

This proof of concept shows the potential of AI-supported clinical tools and at the same time highlights the limitations of using LLMs in medicine. Despite the insufficient correctness and limitations such as model hallucinations and over-reliance on irrelevant patient data, the pediatricians noted the overall value of the AI tab as a decision support tool to assess the growth of children and adolecents especially to get a second opinion in complex cases. Nevertheless, the model's responses need significant improvement in future work, as reliable and correct information are critical in medicine.

---

Folder Structure of the AI Tab

(only new and updated files visible)

.
├── smartapp/
│   ├── ai-tab/
│   │   ├── css
│   │   │   └── gc-ai-summary-view.css
│   │   ├── fixtures
│   │   │   ├── patients
│   │   │   │   ├── ftt_ethan_johnson.json
│   │   │   │   ├── ftt_hannah_becker.json
│   │   │   │   └── ...
│   │   │   ├── createResource.py
│   │   │   ├── deletePatient.py
│   │   │   └── deleteResource.py
│   │   ├── guidelines
│   │   │   ├── UTD Childhood Obesity.pdf
│   │   │   ├── UTD Failure to thrive.pdf
│   │   │   ├── UTD GH Deficiency.pdf
│   │   │   └── UTD T1DM.pdf
│   │   ├── img
│   │   │   ├── chat-button.png
│   │   │   └── ...
│   │   ├── js
│   │   │   ├── openai
│   │   │   │   ├── bundeled-openai.js
│   │   │   │   └── openai-setup.js
│   │   │   ├── gc-ai-summary.js
│   │   │   └── load-patient-history.js
│   │   └── tests
│   │       └── ai-tab.test.js
│   ├── js
│   │   ├── ...
│   │   ├── gc-app.js
│   │   └── ...
│   └── ...
├── ...
├── index.html
├── launch.html
├── load-fhir-data.js
├── README.md
└── ...

---

Overview gc-ai-summary.js

The AI tab module handles various aspects:

• Initializing the AI tab and setting up event listeners.
• Retrieving and processing patient data to generate prompts for growth and history analysis.
• Communicating with the OpenAI assistant for data analysis, recommendations and differential diagnosis.
• Handling user interactions through a chat interface, file uploads, and UI toggling.
• Displaying assistant responses and managing file citations.

Main Functionality Overview:

1. Initialization and Setup

   • Initializes the AI tab, verifies the presence of the OpenAI client, and sets up event listeners.
   • Displays error messages when issues occur during setup or operation.

2. AI Assistant Integration

   • Retrieves the assistant ID from the OpenAI API and sends prompts for analysis.
   • Processes and displays responses, including growth abnormalities, history analysis, specialist referrals, differential diagnoses, and an executive summary.

3. Prompt Generation and Data Processing

   • Fetches patient data and generates prompts for growth and history based on available information.
   • Formats responses from the assistant and updates UI elements accordingly.

4. File Management and Citations

   • Manages file uploads, displays used files, and sets up download buttons for guidelines.
   • Updates content with citation numbers based on file annotations for consistency and traceability.

5. User Interface Handling

   • Handles visibility toggles, chat interactions, and displays popups for growth abnormalities and prompt reviews.
   • Ensures proper formatting and display of data within the AI tab.

6. Patient Data Retrieval and Calculations

   • Retrieves specific patient data such as length, weight, head circumference, BMI, percentiles, Z-scores, and growth velocities.
   • Calculates necessary statistics and formats the information for analysis and presentation.

Structure:

• Initialization: initAITab, showErrorPopup
• AI Interaction: getAssistant, sendToAssistant, createMessage, runAssistant, displayAssistantResponse
• Prompt Creation: fetchPatientData, createPromptGrowth, createPromptHistory, fetchDataAndSendPrompt
• File Handling and Citations: getCitationNumber, updateCitations, getFiles, displayFiles, setupDownloadButtons, initiateDownload
• User Interface Handling: showPopupGrowthWarning, showChat, showPromptPopup, toggleContent
• Data Retrieval and Calculation: getLength, getWeight, getHeadC, getBMI, getPercentile, getZScore, getVelocity, getEntryDate, getDataSet

Deployment

Prerequisite: You must have installed node with npm. Follow this instructions to install node.

0. Run npm:

npm install

1. Before you run the project insert the OpenAI API key I have sent you per email in gc-ai-summary.js (~line 131)

2. To run the project use

npx http-server

3. Login at Logica Health to access the SMART FHIR Server with the synthetic patients:

   3.1 Use the credentials from the email I have sent you:

   • Email: ""
   • Password: ""

   3.2 Click on "thesis-growth-chart-app"

   3.3 Click on "growth-chart-app" - "launch"

   3.4 Choose a physician

   3.5 Choose a patient, e.g. Jacob Miller

   3.6 Click on "growth-chart-app" < "launch"

   3.7 Click "authorize"

---

Tests

The following functions are tested:

• OpenAI initialization: should initialize OpenAI when DOM content is loaded ✓
• getAssistant: should call the getAssistant function and validate the ID ✓
• sendToAssistant: should call sendToAssistant and update the DOM correctly ✓
• getCitationNumber: should assign and retrieve citation numbers correctly ✓
• updateCitations: should correctly update content with citation numbers ✓
• getFiles: should handle errors gracefully and return an empty array ✓
• displayFiles: should update the DOM with new files ✓

1. install the JavaScript testing framework jest (https://jestjs.io)

npm install jest

2. Remove comments at the end of file gc-ai-summary.js before starting the tests

3. Run the tests from the root directory

npx jest

---