app.py

import gradio as gr
import torch
from torch import nn
from transformers import BertTokenizer, BertModel

# Define the BertClassifier class
class BertClassifier(nn.Module):
    def __init__(self, bert: BertModel, num_classes: int):
        super().__init__()
        self.bert = bert
        self.classifier = nn.Linear(bert.config.hidden_size, num_classes)
        self.criterion = nn.BCELoss()

    def forward(self, input_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, labels=None):
        outputs = self.bert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask
        )
        cls_output = outputs.pooler_output
        cls_output = self.classifier(cls_output)
        cls_output = torch.sigmoid(cls_output)

        loss = 0
        if labels is not None:
            loss = self.criterion(cls_output, labels)
        return loss, cls_output

# Load the tokenizer and model
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
bert_model = BertModel.from_pretrained('bert-base-uncased')
model = BertClassifier(bert_model, num_classes=7)

# Load the model weights from the .pkl file
model.load_state_dict(torch.load('bert_classifier_icd.pkl', map_location=torch.device('cpu')))
model.eval()

# Define prediction function
def predict(text):
    tokens = tokenizer.encode(text, add_special_tokens=True, max_length=512, truncation=True)
    input_ids = torch.tensor([tokens])
    mask = (input_ids != tokenizer.pad_token_id).float()
    
    with torch.no_grad():
        _, outputs = model(input_ids, attention_mask=mask)
    
    # Assuming outputs[0] contains the probability scores for each class
    confidence_scores = outputs[0].tolist()
    
    # Convert to a dictionary mapping labels to confidence scores
    labels = ['Cardiovascular', 'Respiratory', 'Neurological', 'Infectious', 'Endocrine', 'Musculoskeletal', 'Gastrointestinal']
    prediction = {label: score for label, score in zip(labels, confidence_scores)}
    
    return prediction

# Add example texts
examples = [
    ["Patient admitted with chest pain, shortness of breath, and abnormal ECG findings."],
    ["Elderly patient presented with symptoms of confusion, fever, and elevated white blood cell count."],
    ["Patient complains of persistent cough, wheezing, and history of asthma."],
    ["Admitted with severe abdominal pain, nausea, and vomiting. Suspected appendicitis."],
    ["Patient has a history of diabetes mellitus and presented with high blood glucose levels and dehydration."],
    ["Patient admitted following a fall, showing signs of fracture in the left femur."],
    ["Patient experiencing severe headaches, dizziness, and a history of epilepsy."],
    ["Acute kidney injury suspected due to elevated creatinine and reduced urine output."],
    ["Patient diagnosed with major depressive disorder, experiencing prolonged sadness and loss of interest in activities."],
    ["Presented with a bacterial skin infection requiring intravenous antibiotics."]
]

# Create Gradio interface
iface = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=10, placeholder="Enter clinical notes here..."),
    outputs=gr.Label(num_top_classes=7),
    examples=examples,
    title="MIMIC-IV ICD Code Classification",
    description="Predict ICD code categories based on clinical text using a BERT-based model. The model outputs confidence scores for seven common ICD code categories.",
)

iface.launch()