app.py

import gradio as gr
import torch
import torch.nn.functional as F
from model import DualEncoderModel
from utils import (
    extract_text_from_pdf,
    extract_lab_tests_dict,
    prepare_lab_tensor,
    load_icd_mapping,
)
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import OpenAIEmbeddings
from langchain.chains import RetrievalQA
from langchain_community.chat_models import ChatOpenAI

lab_cont_features_list = [
    "ALT (SGPT)",
    "AST (SGOT)",
    "Bilirubin",
    "Albumin",
    "Platelet Count",
    "Total Cholesterol",
    "BP Systolic",
    "BP Diastolic",
    "Troponin",
    "Ejection Fraction",
    "HbA1c",
    "Fasting Glucose",
    "Postprandial Glucose",
    "Triglycerides",
    "Insulin Level",
    "WBC Count",
    "Fever",
    "Hematocrit",
]

model_path = "model/dual_encoder_model.pth"
icd_csv_path = "model/augmented_lab_data.csv"

model = DualEncoderModel(
    lab_cont_dim=len(lab_cont_features_list),
    lab_cat_dims=[],
    conv_cont_dim=0,
    conv_cat_dims=[49, 17, 17],
    embedding_dim=16,
    num_classes=18,
)
model.load_state_dict(torch.load(model_path, map_location=torch.device("cpu")))
model.eval()

icd_mapping = load_icd_mapping(icd_csv_path)


def predict_icd(pdf):
    text = extract_text_from_pdf(pdf.name)
    splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
    docs = splitter.create_documents([text])

    embedding_model = OpenAIEmbeddings(model="text-embedding-3-large")
    vectorstore = FAISS.from_documents(docs, embedding=embedding_model)
    retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
    llm = ChatOpenAI(model_name="gpt-4o", temperature=0)
    qa = RetrievalQA.from_chain_type(llm=llm, retriever=retriever, chain_type="refine")

    query = "List lab test names and values only with units (no suggestions). Format: Test: Value Unit"
    lab_response = qa.run(query)
    lab_data = extract_lab_tests_dict(lab_response)

    lab_cont_tensor = prepare_lab_tensor(lab_data, lab_cont_features_list)
    lab_cat_tensor = torch.zeros((1, 0), dtype=torch.int64)
    conv_cont_tensor = torch.zeros((1, 0), dtype=torch.float32)
    conv_cat_tensor = torch.tensor([[0, 0, 0]], dtype=torch.int64)

    with torch.no_grad():
        logits = model(
            lab_cont_tensor, lab_cat_tensor, conv_cont_tensor, conv_cat_tensor
        )
        probs = F.softmax(logits, dim=1)
        top_probs = torch.topk(probs, 3, dim=1)

    output = ""
    for i, (label_idx, prob) in enumerate(
        zip(top_probs.indices[0], top_probs.values[0])
    ):
        icd_code, icd_label, diagnosis = icd_mapping.get(
            label_idx.item(), ("Unknown", "Unknown", "No Description Available")
        )
        confidence = (
            "🔵 High" if prob > 0.6 else "🟡 Medium" if prob > 0.3 else "🔴 Low"
        )
        output += f"{i+1}. **{diagnosis}**\nICD Code: {icd_code}\nConfidence: {confidence} ({prob:.2%})\n\n"
    return output.strip()


iface = gr.Interface(
    fn=predict_icd,
    inputs=gr.File(label="Upload PDF Lab Report"),
    outputs=gr.Markdown(label="Predicted Diagnoses (ICD Codes)"),
    title="ICD Code Predictor from Lab Report",
    description="Upload a lab report PDF to predict possible diagnoses with ICD codes.",
)

if __name__ == "__main__":
    iface.launch()