app.py

import pandas as pd
import json
import gradio as gr
import numpy as np
import torch
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import classification_report
from sklearn.preprocessing import LabelEncoder
from transformers import AutoTokenizer, AutoModel

# Load datasets from local storage
train_df = pd.read_csv("Train_dataset.csv")
test_df = pd.read_csv("Test_dataset.csv")

# Load disease mapping
with open("disease_mapping.json", "r") as f:
    disease_info = {item["Disease"]: item for item in json.load(f)}

# Encode disease labels
le = LabelEncoder()
train_df['label'] = le.fit_transform(train_df['Disease'])

# Filter out test samples with unseen diseases
test_df = test_df[test_df['Disease'].isin(le.classes_)]
test_df['label'] = le.transform(test_df['Disease'])

# Load SciBERT tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("allenai/scibert_scivocab_uncased")
model = AutoModel.from_pretrained("allenai/scibert_scivocab_uncased")

# Function to get [CLS] token embedding
def get_embedding(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding='max_length', max_length=128)
    with torch.no_grad():
        outputs = model(**inputs)
    return outputs.last_hidden_state[:, 0, :].squeeze().numpy()

# Generate embeddings for training and testing
print("Generating embeddings for training data...")
X_train = np.vstack([get_embedding(text) for text in train_df['text']])
y_train = train_df['label'].values

print("Generating embeddings for test data...")
X_test = np.vstack([get_embedding(text) for text in test_df['text']])
y_test = test_df['label'].values

# Train Gradient Boosting classifier
print("Training classifier...")
clf = GradientBoostingClassifier()
clf.fit(X_train, y_train)

# Prediction function
def predict_disease(symptoms):
    emb = get_embedding(symptoms).reshape(1, -1)
    probs = clf.predict_proba(emb)[0]
    top3_idx = np.argsort(probs)[::-1][:3]
    results = []
    for idx in top3_idx:
        disease = le.inverse_transform([idx])[0]
        info = disease_info.get(disease, {})
        results.append({
            "Disease": disease,
            "Confidence": round(probs[idx] * 100, 2),
            "Description": info.get("Description", "N/A"),
            "Severity": info.get("Severity", "N/A"),
            "Precaution": info.get("Precaution", "N/A")
        })
    return results

# Gradio chatbot interface
def chatbot_interface(symptom_text):
    preds = predict_disease(symptom_text)
    output = ""
    for i, pred in enumerate(preds, 1):
        output += f"### Prediction {i}\n"
        output += f"- **Disease:** {pred['Disease']} ({pred['Confidence']}%)\n"
        output += f"- **Description:** {pred['Description']}\n"
        output += f"- **Severity:** {pred['Severity']}\n"
        output += f"- **Precaution:** {pred['Precaution']}\n\n"
    return output.strip()

# Launch Gradio UI
gr.Interface(
    fn=chatbot_interface,
    inputs=gr.Textbox(label="Enter your symptoms"),
    outputs=gr.Markdown(),
    title="SciBERT Medical Chatbot",
    description="AI Medical Assistant that predicts diseases based on symptoms using SciBERT embeddings."
).launch()