icd9_ui.py

# import streamlit as st
# import torch
# from transformers import LongformerTokenizer, LongformerForSequenceClassification

# # Load the fine-tuned model and tokenizer
# model_path = "./clinical_longformer"
# tokenizer = LongformerTokenizer.from_pretrained(model_path)
# model = LongformerForSequenceClassification.from_pretrained(model_path)
# model.eval()  # Set the model to evaluation mode

# # ICD-9 code columns used during training
# icd9_columns = [
#     '038.9', '244.9', '250.00', '272.0', '272.4', '276.1', '276.2', '285.1', '285.9',
#     '287.5', '305.1', '311', '36.15', '37.22', '37.23', '38.91', '38.93', '39.61',
#     '39.95', '401.9', '403.90', '410.71', '412', '414.01', '424.0', '427.31', '428.0',
#     '486', '496', '507.0', '511.9', '518.81', '530.81', '584.9', '585.9', '599.0',
#     '88.56', '88.72', '93.90', '96.04', '96.6', '96.71', '96.72', '99.04', '99.15',
#     '995.92', 'V15.82', 'V45.81', 'V45.82', 'V58.61'
# ]

# # Function for making predictions
# def predict_icd9(texts, tokenizer, model, threshold=0.5):
#     inputs = tokenizer(
#         texts,
#         padding="max_length",
#         truncation=True,
#         max_length=512,
#         return_tensors="pt"
#     )
    
#     with torch.no_grad():
#         outputs = model(
#             input_ids=inputs["input_ids"],
#             attention_mask=inputs["attention_mask"]
#         )
#         logits = outputs.logits
#         probabilities = torch.sigmoid(logits)
#         predictions = (probabilities > threshold).int()
    
#     predicted_icd9 = []
#     for pred in predictions:
#         codes = [icd9_columns[i] for i, val in enumerate(pred) if val == 1]
#         predicted_icd9.append(codes)
    
#     return predicted_icd9

# # Streamlit UI
# st.title("ICD-9 Code Prediction")
# st.sidebar.header("Model Options")
# model_option = st.sidebar.selectbox("Select Model", [ "ClinicalLongformer"])
# threshold = st.sidebar.slider("Prediction Threshold", 0.0, 1.0, 0.5, 0.01)

# st.write("### Enter Medical Summary")
# input_text = st.text_area("Medical Summary", placeholder="Enter clinical notes here...")

# if st.button("Predict"):
#     if input_text.strip():
#         predictions = predict_icd9([input_text], tokenizer, model, threshold)
#         st.write("### Predicted ICD-9 Codes")
#         for code in predictions[0]:
#             st.write(f"- {code}")
#     else:
#         st.error("Please enter a medical summary.")

# import torch
# import pandas as pd
# import streamlit as st
# from transformers import LongformerTokenizer, LongformerForSequenceClassification

# # Load the fine-tuned model and tokenizer
# model_path = "./clinical_longformer"
# tokenizer = LongformerTokenizer.from_pretrained(model_path)
# model = LongformerForSequenceClassification.from_pretrained(model_path)
# model.eval()  # Set the model to evaluation mode

# # Load the ICD-9 descriptions from CSV into a dictionary
# icd9_desc_df = pd.read_csv("D_ICD_DIAGNOSES.csv")  # Adjust the path to your CSV file
# icd9_desc_df['ICD9_CODE'] = icd9_desc_df['ICD9_CODE'].astype(str)  # Ensure ICD9_CODE is string type for matching
# icd9_descriptions = dict(zip(icd9_desc_df['ICD9_CODE'].str.replace('.', ''), icd9_desc_df['LONG_TITLE']))  # Remove decimals in ICD9 code for matching

# # ICD-9 code columns used during training
# icd9_columns = [
#     '038.9', '244.9', '250.00', '272.0', '272.4', '276.1', '276.2', '285.1', '285.9',
#     '287.5', '305.1', '311', '36.15', '37.22', '37.23', '38.91', '38.93', '39.61',
#     '39.95', '401.9', '403.90', '410.71', '412', '414.01', '424.0', '427.31', '428.0',
#     '486', '496', '507.0', '511.9', '518.81', '530.81', '584.9', '585.9', '599.0',
#     '88.56', '88.72', '93.90', '96.04', '96.6', '96.71', '96.72', '99.04', '99.15',
#     '995.92', 'V15.82', 'V45.81', 'V45.82', 'V58.61'
# ]

# # Function for making predictions
# def predict_icd9(texts, tokenizer, model, threshold=0.5):
#     inputs = tokenizer(
#         texts,
#         padding="max_length",
#         truncation=True,
#         max_length=512,
#         return_tensors="pt"
#     )
    
#     with torch.no_grad():
#         outputs = model(
#             input_ids=inputs["input_ids"],
#             attention_mask=inputs["attention_mask"]
#         )
#         logits = outputs.logits
#         probabilities = torch.sigmoid(logits)
#         predictions = (probabilities > threshold).int()
    
#     predicted_icd9 = []
#     for pred in predictions:
#         codes = [icd9_columns[i] for i, val in enumerate(pred) if val == 1]
#         predicted_icd9.append(codes)
    
#     # Fetch descriptions for the predicted ICD-9 codes from the pre-loaded descriptions
#     predictions_with_desc = []
#     for codes in predicted_icd9:
#         code_with_desc = [(code, icd9_descriptions.get(code.replace('.', ''), "Description not found")) for code in codes]
#         predictions_with_desc.append(code_with_desc)
    
#     return predictions_with_desc

# # Streamlit UI
# st.title("ICD-9 Code Prediction")
# st.sidebar.header("Model Options")
# threshold = st.sidebar.slider("Prediction Threshold", 0.0, 1.0, 0.5, 0.01)

# st.write("### Enter Medical Summary")
# input_text = st.text_area("Medical Summary", placeholder="Enter clinical notes here...")

# if st.button("Predict"):
#     if input_text.strip():
#         predictions = predict_icd9([input_text], tokenizer, model, threshold)
#         st.write("### Predicted ICD-9 Codes and Descriptions")
#         for code, description in predictions[0]:
#             st.write(f"- {code}: {description}")
#     else:
#         st.error("Please enter a medical summary.")
# import torch
# import pandas as pd
# import streamlit as st
# from transformers import LongformerTokenizer, LongformerForSequenceClassification

# # Load the fine-tuned model and tokenizer
# model_path = "./clinical_longformer"
# tokenizer = LongformerTokenizer.from_pretrained(model_path)
# model = LongformerForSequenceClassification.from_pretrained(model_path)
# model.eval()  # Set the model to evaluation mode

# # Load the ICD-9 descriptions from CSV into a dictionary
# icd9_desc_df = pd.read_csv("D_ICD_DIAGNOSES.csv")  # Adjust the path to your CSV file
# icd9_desc_df['ICD9_CODE'] = icd9_desc_df['ICD9_CODE'].astype(str)  # Ensure ICD9_CODE is string type
# icd9_descriptions = dict(zip(icd9_desc_df['ICD9_CODE'].str.replace('.', ''), icd9_desc_df['LONG_TITLE']))  # Remove decimals for matching

# # Load the ICD-9 to ICD-10 mapping
# icd9_to_icd10 = {}
# with open("2015_I9gem.txt", "r") as file:
#     for line in file:
#         parts = line.strip().split()
#         if len(parts) == 3:
#             icd9, icd10, _ = parts
#             icd9_to_icd10[icd9] = icd10

# # ICD-9 code columns used during training
# icd9_columns = [
#     '038.9', '244.9', '250.00', '272.0', '272.4', '276.1', '276.2', '285.1', '285.9',
#     '287.5', '305.1', '311', '36.15', '37.22', '37.23', '38.91', '38.93', '39.61',
#     '39.95', '401.9', '403.90', '410.71', '412', '414.01', '424.0', '427.31', '428.0',
#     '486', '496', '507.0', '511.9', '518.81', '530.81', '584.9', '585.9', '599.0',
#     '88.56', '88.72', '93.90', '96.04', '96.6', '96.71', '96.72', '99.04', '99.15',
#     '995.92', 'V15.82', 'V45.81', 'V45.82', 'V58.61'
# ]

# # Function for making predictions and mapping to ICD-10
# def predict_icd9(texts, tokenizer, model, threshold=0.5):
#     inputs = tokenizer(
#         texts,
#         padding="max_length",
#         truncation=True,
#         max_length=512,
#         return_tensors="pt"
#     )

#     with torch.no_grad():
#         outputs = model(
#             input_ids=inputs["input_ids"],
#             attention_mask=inputs["attention_mask"]
#         )
#         logits = outputs.logits
#         probabilities = torch.sigmoid(logits)
#         predictions = (probabilities > threshold).int()

#     predicted_icd9 = []
#     for pred in predictions:
#         codes = [icd9_columns[i] for i, val in enumerate(pred) if val == 1]
#         predicted_icd9.append(codes)

#     # Fetch descriptions and map to ICD-10 codes
#     predictions_with_desc = []
#     for codes in predicted_icd9:
#         code_with_desc = []
#         for code in codes:
#             icd9_stripped = code.replace('.', '')
#             icd10_code = icd9_to_icd10.get(icd9_stripped, "Mapping not found")
#             icd9_desc = icd9_descriptions.get(icd9_stripped, "Description not found")
#             code_with_desc.append((code, icd9_desc, icd10_code))
#         predictions_with_desc.append(code_with_desc)

#     return predictions_with_desc

# # Streamlit UI
# st.title("ICD-9 to ICD-10 Code Prediction")
# st.sidebar.header("Model Options")
# threshold = st.sidebar.slider("Prediction Threshold", 0.0, 1.0, 0.5, 0.01)

# st.write("### Enter Medical Summary")
# input_text = st.text_area("Medical Summary", placeholder="Enter clinical notes here...")

# if st.button("Predict"):
#     if input_text.strip():
#         predictions = predict_icd9([input_text], tokenizer, model, threshold)
#         st.write("### Predicted ICD-9 and ICD-10 Codes with Descriptions")
#         for icd9_code, description, icd10_code in predictions[0]:
#             st.write(f"- ICD-9: {icd9_code} ({description}) -> ICD-10: {icd10_code}")
#     else:
#         st.error("Please enter a medical summary.")

import os
import torch
import pandas as pd
import streamlit as st
from PIL import Image
from transformers import LongformerTokenizer, LongformerForSequenceClassification
from phi.agent import Agent
from phi.model.google import Gemini
from phi.tools.duckduckgo import DuckDuckGo

# Load the fine-tuned ICD-9 model and tokenizer
model_path = "./clinical_longformer"
tokenizer = LongformerTokenizer.from_pretrained(model_path)
model = LongformerForSequenceClassification.from_pretrained(model_path)
model.eval()  # Set the model to evaluation mode

# Load the ICD-9 descriptions from CSV into a dictionary
icd9_desc_df = pd.read_csv("D_ICD_DIAGNOSES.csv")  # Adjust the path to your CSV file
icd9_desc_df['ICD9_CODE'] = icd9_desc_df['ICD9_CODE'].astype(str)  # Ensure ICD9_CODE is string type for matching
icd9_descriptions = dict(zip(icd9_desc_df['ICD9_CODE'].str.replace('.', ''), icd9_desc_df['LONG_TITLE']))  # Remove decimals in ICD9 code for matching

# ICD-9 code columns used during training
icd9_columns = [
    '038.9', '244.9', '250.00', '272.0', '272.4', '276.1', '276.2', '285.1', '285.9',
    '287.5', '305.1', '311', '36.15', '37.22', '37.23', '38.91', '38.93', '39.61',
    '39.95', '401.9', '403.90', '410.71', '412', '414.01', '424.0', '427.31', '428.0',
    '486', '496', '507.0', '511.9', '518.81', '530.81', '584.9', '585.9', '599.0',
    '88.56', '88.72', '93.90', '96.04', '96.6', '96.71', '96.72', '99.04', '99.15',
    '995.92', 'V15.82', 'V45.81', 'V45.82', 'V58.61'
]

# Function for making ICD-9 predictions
def predict_icd9(texts, tokenizer, model, threshold=0.5):
    inputs = tokenizer(
        texts,
        padding="max_length",
        truncation=True,
        max_length=512,
        return_tensors="pt"
    )
    
    with torch.no_grad():
        outputs = model(
            input_ids=inputs["input_ids"],
            attention_mask=inputs["attention_mask"]
        )
        logits = outputs.logits
        probabilities = torch.sigmoid(logits)
        predictions = (probabilities > threshold).int()
    
    predicted_icd9 = []
    for pred in predictions:
        codes = [icd9_columns[i] for i, val in enumerate(pred) if val == 1]
        predicted_icd9.append(codes)
    
    predictions_with_desc = []
    for codes in predicted_icd9:
        code_with_desc = [(code, icd9_descriptions.get(code.replace('.', ''), "Description not found")) for code in codes]
        predictions_with_desc.append(code_with_desc)
    
    return predictions_with_desc

# Streamlit UI
st.title("Medical Diagnosis Assistant")
option = st.selectbox(
    "Choose Diagnosis Method",
    ("ICD-9 Code Prediction", "Medical Image Analysis")
)

# ICD-9 Code Prediction
if option == "ICD-9 Code Prediction":
    st.write("### Enter Medical Summary")
    input_text = st.text_area("Medical Summary", placeholder="Enter clinical notes here...")

    threshold = st.slider("Prediction Threshold", 0.0, 1.0, 0.5, 0.01)

    if st.button("Predict ICD-9 Codes"):
        if input_text.strip():
            predictions = predict_icd9([input_text], tokenizer, model, threshold)
            st.write("### Predicted ICD-9 Codes and Descriptions")
            for code, description in predictions[0]:
                st.write(f"- {code}: {description}")
        else:
            st.error("Please enter a medical summary.")

# Medical Image Analysis
elif option == "Medical Image Analysis":
    if "GOOGLE_API_KEY" not in st.session_state:
        st.warning("Please enter your Google API Key in the sidebar to continue")
    else:
        medical_agent = Agent(
            model=Gemini(
                api_key=st.session_state.GOOGLE_API_KEY,
                id="gemini-2.0-flash-exp"
            ),
            tools=[DuckDuckGo()],
            markdown=True
        )

        query = """
        You are a highly skilled medical imaging expert with extensive knowledge in radiology and diagnostic imaging. Analyze the patient's medical image and structure your response as follows:

        ### 1. Image Type & Region
        - Specify imaging modality (X-ray/MRI/CT/Ultrasound/etc.)
        - Identify the patient's anatomical region and positioning
        - Comment on image quality and technical adequacy

        ### 2. Key Findings
        - List primary observations systematically
        - Note any abnormalities in the patient's imaging with precise descriptions
        - Include measurements and densities where relevant
        - Describe location, size, shape, and characteristics
        - Rate severity: Normal/Mild/Moderate/Severe

        ### 3. Diagnostic Assessment
        - Provide primary diagnosis with confidence level
        - List differential diagnoses in order of likelihood
        - Support each diagnosis with observed evidence from the patient's imaging
        - Note any critical or urgent findings

        ### 4. Patient-Friendly Explanation
        - Explain the findings in simple, clear language that the patient can understand
        - Avoid medical jargon or provide clear definitions
        - Include visual analogies if helpful
        - Address common patient concerns related to these findings

        ### 5. Research Context
        - Use the DuckDuckGo search tool to find recent medical literature about similar cases
        - Provide a list of relevant medical links
        - Include key references to support your analysis
        """

        upload_container = st.container()
        image_container = st.container()
        analysis_container = st.container()

        with upload_container:
            uploaded_file = st.file_uploader(
                "Upload Medical Image",
                type=["jpg", "jpeg", "png", "dicom"],
                help="Supported formats: JPG, JPEG, PNG, DICOM"
            )

        if uploaded_file is not None:
            with image_container:
                col1, col2, col3 = st.columns([1, 2, 1])
                with col2:
                    image = Image.open(uploaded_file)
                    width, height = image.size
                    aspect_ratio = width / height
                    new_width = 500
                    new_height = int(new_width / aspect_ratio)
                    resized_image = image.resize((new_width, new_height))
                    
                    st.image(resized_image, caption="Uploaded Medical Image", use_container_width=True)
                    
                    analyze_button = st.button("🔍 Analyze Image")

            with analysis_container:
                if analyze_button:
                    image_path = "temp_medical_image.png"
                    with open(image_path, "wb") as f:
                        f.write(uploaded_file.getbuffer())
                    
                    with st.spinner("🔄 Analyzing image... Please wait."):
                        try:
                            response = medical_agent.run(query, images=[image_path])
                            st.markdown("### 📋 Analysis Results")
                            st.markdown(response.content)
                        except Exception as e:
                            st.error(f"Analysis error: {e}")
                        finally:
                            if os.path.exists(image_path):
                                os.remove(image_path)
        else:
            st.info("👆 Please upload a medical image to begin analysis")