app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModel
import torch.nn.functional as F
import os

# Define the model class
class MedicalCodePredictor(torch.nn.Module):
    def __init__(self, bert_model):
        super().__init__()
        self.bert = bert_model
        self.dropout = torch.nn.Dropout(0.1)
        self.icd_classifier = torch.nn.Linear(768, len(ICD_CODES))
        self.cpt_classifier = torch.nn.Linear(768, len(CPT_CODES))
        
    def forward(self, input_ids, attention_mask):
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = outputs.last_hidden_state[:, 0, :]
        pooled_output = self.dropout(pooled_output)
        
        icd_logits = self.icd_classifier(pooled_output)
        cpt_logits = self.cpt_classifier(pooled_output)
        
        return icd_logits, cpt_logits

# Load ICD codes from files
def load_icd_codes_from_files():
    icd_codes = {}
    directory_path = "./codes/icd_txt_files/"  # Path to ICD codes directory

    if os.path.exists(directory_path):
        for file_name in os.listdir(directory_path):
            if file_name.endswith(".txt"):
                file_path = os.path.join(directory_path, file_name)
                with open(file_path, "r", encoding="utf-8") as file:
                    for line in file:
                        # Skip empty lines
                        if line.strip():
                            # Split the line into code and description
                            parts = line.strip().split(maxsplit=1)
                            if len(parts) == 2:
                                code = parts[0].strip()
                                description = parts[1].strip()
                                icd_codes[code] = description
                            else:
                                print(f"Invalid line format in file {file_name}: {line}")
    else:
        print(f"Directory {directory_path} does not exist!")

    if not icd_codes:
        raise ValueError("No ICD codes were loaded. Please check your files and directory structure.")

    return icd_codes

    ICD_CODES = load_icd_codes_from_files()
    print(f"Loaded {len(ICD_CODES)} ICD codes.")

# Load CPT codes from files
def load_cpt_codes_from_files():
    cpt_codes = {}
    directory_path = "./codes/cpt_txt_files/"  # Path to CPT codes directory

    if os.path.exists(directory_path):
        for file_name in os.listdir(directory_path):
            if file_name.endswith(".txt"):
                file_path = os.path.join(directory_path, file_name)
                with open(file_path, "r", encoding="utf-8") as file:
                    for line in file:
                        # Split the line into code and description
                        parts = line.strip().split(maxsplit=1)
                        if len(parts) == 2:
                            code = parts[0].strip()
                            description = parts[1].strip()
                            cpt_codes[code] = description
    else:
        print(f"Directory {directory_path} does not exist!")

    return cpt_codes

# Load ICD and CPT codes dynamically
ICD_CODES = load_icd_codes_from_files()
CPT_CODES = load_cpt_codes_from_files()

# Load models
@torch.no_grad()
def load_models():
    tokenizer = AutoTokenizer.from_pretrained("emilyalsentzer/Bio_ClinicalBERT")
    base_model = AutoModel.from_pretrained("emilyalsentzer/Bio_ClinicalBERT")
    model = MedicalCodePredictor(base_model)
    return tokenizer, model

# Prediction function
def predict_codes(text):
    if not text.strip():
        return "Please enter a medical summary."
    
    # Tokenize input
    inputs = tokenizer(text, 
                      return_tensors="pt",
                      max_length=512,
                      truncation=True,
                      padding=True)
    
    # Get predictions
    model.eval()
    icd_logits, cpt_logits = model(inputs['input_ids'], inputs['attention_mask'])
    
    # Get probabilities
    icd_probs = F.softmax(icd_logits, dim=1)
    cpt_probs = F.softmax(cpt_logits, dim=1)
    
    # Get top 3 predictions
    top_icd = torch.topk(icd_probs, k=3)
    top_cpt = torch.topk(cpt_probs, k=3)

    # Get top k predictions (limit k to the number of available codes)
    top_k = min(3, len(ICD_CODES))
    top_icd = torch.topk(icd_probs, k=top_k)

    
    # Format results
    result = "Recommended ICD-10 Codes:\n"
    for i, (prob, idx) in enumerate(zip(top_icd.values[0], top_icd.indices[0])):
        result += f"{i+1}. {ICD_CODES.get(idx.item(), 'Unknown')} (Confidence: {prob.item():.2f})\n"
    
    result += "\nRecommended CPT Codes:\n"
    for i, (prob, idx) in enumerate(zip(top_cpt.values[0], top_cpt.indices[0])):
        result += f"{i+1}. {CPT_CODES.get(idx.item(), 'Unknown')} (Confidence: {prob.item():.2f})\n"
    
    return result

# Load models globally
tokenizer, model = load_models()

# Create Gradio interface
iface = gr.Interface(
    fn=predict_codes,
    inputs=gr.Textbox(
        lines=5,
        placeholder="Enter medical summary here...",
        label="Medical Summary"
    ),
    outputs=gr.Textbox(
        label="Predicted Codes",
        lines=8
    ),
    title="AutoRCM - Medical Code Predictor",
    description="Enter a medical summary to get recommended ICD-10 and CPT codes.",
    examples=[
        ["Patient presents with blood pressure 150/90. Complains of occasional headaches. History of hypertension."],
        ["Patient has elevated blood sugar levels. A1C is 7.8. History of type 2 diabetes."],
        ["Patient complains of chronic lower back pain, worse with movement. No radiation to legs."]
    ]
)

# Launch the interface
iface.launch(share=True)