app.py

import streamlit as st
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import pandas as pd
import numpy as np
from sklearn.preprocessing import MultiLabelBinarizer

# Check if a GPU is available
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load the trained model and tokenizer

@st.cache_resource
def load_model():
    model = AutoModelForSequenceClassification.from_pretrained(
        "microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract",
        num_labels=8,  # Adjust based on your label count
        problem_type="multi_label_classification"
    )
    # Map the model to the appropriate device
    model.load_state_dict(torch.load('best_model_v2.pth', map_location=torch.device('cpu')))
    model.eval()
    tokenizer = AutoTokenizer.from_pretrained("microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract")
    return model, tokenizer


@st.cache_resource
def load_mlb():
    # Define the classes based on your label set
    classes = ['81001.0','99213.0','99214.0','E11.9','I10','J45.909','M54.5','N39.0']
    # Initialize and fit the MultiLabelBinarizer
    mlb = MultiLabelBinarizer(classes=classes)
    mlb.fit([classes])  # Fit with the full list of labels as a single sample
    
    return mlb


# # Load MultiLabelBinarizer
# @st.cache_resource
# def load_mlb():
#     mlb = MultiLabelBinarizer()
#     # mlb.classes_ = np.load('mlb_classes.npy')  # Assuming you saved the classes array during training
#     mlb = MultiLabelBinarizer(classes=['E11.9', 'I10', 'J45.909', 'M54.5',
#        'N39.0', '81001.0', '99213.0', '99214.0'])  # Update with actual labels

#     return mlb

model, tokenizer = load_model()
mlb = load_mlb()

# Streamlit UI
st.title("Automated Medical Coding")
# st.write("Enter clinical notes to predict ICD and CPT codes.")

# Text input for Clinical Notes
clinical_note = st.text_area("Enter clinical notes")

# Prediction button
if st.button('Predict'):
    if clinical_note:
        # Tokenize the input clinical note
        inputs = tokenizer(clinical_note, truncation=True, padding="max_length", max_length=512, return_tensors='pt')
        
        # Move inputs to the GPU if available
        # inputs = {key: val.to(device) for key, val in inputs.items()}
        inputs = {key: val.to(torch.device('cpu')) for key, val in inputs.items()}

        
        # Model inference
        with torch.no_grad():
            outputs = model(**inputs)
            logits = outputs.logits
        
        # Apply sigmoid and threshold the output (0.5 for multi-label classification)
        pred_labels = (torch.sigmoid(logits) > 0.5).cpu().numpy()
        
        # Get the predicted ICD and CPT codes
        predicted_codes = mlb.inverse_transform(pred_labels)
        
        # Format the results for better display
        if predicted_codes:
            st.write("**Predicted CPT and ICD Codes:**")
            for codes in predicted_codes:
                for code in codes:
                    if code in ['81001.0', '99213.0', '99214.0']:  # Adjust based on your CPT code list
                        st.write(f"- **CPT Code:** {code}")
                    else:
                        st.write(f"- **ICD Code:** {code}")
        else:
            st.write("No codes predicted.")
        
    else:
        st.write("Please enter clinical notes for prediction.")