Update app.py

app.py

@@ -1,718 +1,571 @@
-# -*- coding: utf-8 
+# -*- coding: utf-8 -*-
+# 🏥 Gemma 3N SOAP Note Generator with Unsloth
+# Optimized for offline medical documentation
 
-# 🏥 Gemma 3N SOAP Note Generator
-#-*- coding: utf-8 
-
-# 🏥 Gemma 3N SOAP Note Generator
-
-
-# Enable widgets
-
-
-
-# Enable widgets
-
-
-import torch
-from transformers import AutoProcessor, AutoModelForImageTextToText
-import gradio as gr
-import ipywidgets as widgets
-from IPython.display import display, clear_output
-import io
-import base64
-from datetime import datetime
-from huggingface_hub import login
-import getpass
-
-# Authenticate with HuggingFace
-# Replace the authentication section (lines around the getpass part) with this:
-
-# Import libraries and authenticate
 import torch
-from transformers import AutoProcessor, AutoModelForImageTextToText
 import gradio as gr
-import ipywidgets as widgets
-from IPython.display import display, clear_output
 import io
 import base64
 from datetime import datetime
-from huggingface_hub import login
 import os
 import easyocr
-from PIL import Image
-
-
-
-# Authenticate with HuggingFace
-print("🔐 HuggingFace Authentication Required")
-
-# Try to get token from environment variable first (for production/HF Spaces)
-hf_token = os.environ.get('HF_TOKEN') or os.environ.get('HUGGINGFACE_TOKEN')
+from PIL import Image, ImageDraw, ImageFont
+import cv2
+import numpy as np
+import psutil
 
-if hf_token:
-    print("✅ Found HF token in environment variables")
+# Import Unsloth for optimized Gemma 3n
 try:
-    login(token=hf_token)
-    print("✅ Successfully authenticated with HuggingFace!")
-except Exception as e:
-    print(f"❌ Authentication failed: {e}")
-    print("Please check your token and try again.")
-# Check GPU availability
-import os
-os.environ["CUDA_VISIBLE_DEVICES"] = ""   # Hide all GPUs from PyTorch
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-device = "cpu"
-print(f"🖥️  Using device: {device}")
-if torch.cuda.is_available():
-    print(f"🚀 GPU: {torch.cuda.get_device_name(0)}")
-    print(f"💾 GPU Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
-else:
-    print("⚠️  Running on CPU - this will be slower")
-
-# Load Gemma 3N model
-print("📡 Loading Gemma 3N model...")
-model_id = "google/gemma-3n-e2b-it"
-
-print("🔧 Loading processor...")
-processor = AutoProcessor.from_pretrained(model_id)
-
-print("🤖 Loading Gemma 3N model (this may take a few minutes)...")
-model = AutoModelForImageTextToText.from_pretrained(
-    model_id,
-    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
-    low_cpu_mem_usage=True,
-).to(device)
+    from unsloth import FastModel
+    print("✅ Unsloth imported successfully")
+    UNSLOTH_AVAILABLE = True
+except ImportError:
+    print("❌ Unsloth not available. Install with: pip install unsloth")
+    UNSLOTH_AVAILABLE = False
+
+# Device setup
+def setup_device():
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"🖥️  Using device: {device}")
+    
+    if torch.cuda.is_available():
+        print(f"🚀 GPU: {torch.cuda.get_device_name(0)}")
+        print(f"💾 GPU Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
+    else:
+        print("⚠️  Running on CPU - will be slower but works offline")
+    
+    return device
+
+# Load Unsloth Gemma 3n model
+def load_unsloth_gemma_model(device):
+    """Load optimized Gemma 3n model using Unsloth"""
+    
+    if not UNSLOTH_AVAILABLE:
+        print("❌ Unsloth not available. Using fallback method.")
+        return load_fallback_model()
+    
+    try:
+        print("📡 Loading Unsloth-optimized Gemma 3n model...")
+        
+        # Use the 4-bit quantized model for efficiency
+        model_name = "unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit"
+        
+        print(f"🔧 Loading model: {model_name}")
+        
+        # Load with Unsloth optimizations
+        model, tokenizer = FastModel.from_pretrained(
+            model_name=model_name,
+            dtype=None,  # Auto-detect
+            max_seq_length=1024,  # Good for medical notes
+            load_in_4bit=True,   # 4-bit quantization for efficiency
+            full_finetuning=False,
+        )
+        
+        print("✅ Unsloth Gemma 3n model loaded successfully!")
+        print(f"📊 Model: {model_name}")
+        print(f"💾 Memory optimized with 4-bit quantization")
+        print(f"🎯 Ready for medical SOAP note generation!")
+        
+        return model, tokenizer
+        
+    except Exception as e:
+        print(f"❌ Error loading Unsloth model: {e}")
+        print("💡 Trying fallback model...")
+        return load_fallback_model()
 
-print("✅ Gemma 3N model loaded successfully!")
-print(f"📊 Model size: ~2.9GB")
-print(f"🎯 Ready for SOAP note generation!")
+def load_fallback_model():
+    """Fallback model if Unsloth fails"""
+    try:
+        from transformers import AutoTokenizer, AutoModelForCausalLM
+        
+        print("🔄 Loading fallback model...")
+        model_name = "microsoft/DialoGPT-medium"
+        
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        if tokenizer.pad_token is None:
+            tokenizer.pad_token = tokenizer.eos_token
+            
+        model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+            low_cpu_mem_usage=True
+        )
+        
+        print("✅ Fallback model loaded!")
+        return model, tokenizer
+        
+    except Exception as e:
+        print(f"❌ Fallback model also failed: {e}")
+        return None, None
 
-# SOAP Note Generation Function
-def generate_soap_note(doctor_notes, include_timestamp=True):
-    """
-    Generate a SOAP note from unstructured doctor's notes
-    """
+# Enhanced SOAP Note Generation with Gemma 3n
+def generate_soap_note_gemma(doctor_notes, model=None, tokenizer=None, include_timestamp=True):
+    """Generate SOAP note using Gemma 3n model"""
+    
     if not doctor_notes.strip():
         return "❌ Please enter some medical notes to process."
-
-    prompt = f"""You are a medical AI assistant. Convert the following unstructured doctor's notes into a professional SOAP note format.
-
-Doctor's Notes:
+    
+    if model is None or tokenizer is None:
+        return generate_template_soap(doctor_notes, include_timestamp)
+    
+    # Medical-specific prompt for Gemma 3n
+    prompt = f"""<bos><start_of_turn>user
+You are a medical AI assistant specialized in creating SOAP notes. Convert the following unstructured medical notes into a professional SOAP note format.
+
+Medical Notes:
 {doctor_notes}
 
-Please generate a structured SOAP note with the following sections:
-- SUBJECTIVE: Patient's reported symptoms and history
-- OBJECTIVE: Physical examination findings, vital signs, and test results
-- ASSESSMENT: Clinical diagnosis and reasoning
-- PLAN: Treatment plan, medications, and follow-up
+Please create a structured SOAP note with these sections:
+- SUBJECTIVE: Patient's reported symptoms, complaints, and relevant history
+- OBJECTIVE: Physical examination findings, vital signs, and observable data  
+- ASSESSMENT: Clinical diagnosis, differential diagnosis, and medical reasoning
+- PLAN: Treatment recommendations, medications, tests, and follow-up care
 
-Format your response as a proper medical SOAP note with specific details extracted from the notes."""
+<end_of_turn>
+<start_of_turn>model
+SOAP NOTE:
 
-    try:
-        # Process input
-        inputs = processor(text=prompt, return_tensors="pt").to(device)
+SUBJECTIVE:"""
 
-        # Generate response
-        print("🔄 Generating SOAP note with Gemma 3N...")
+    try:
+        # Tokenize input
+        inputs = tokenizer(
+            prompt, 
+            return_tensors="pt", 
+            truncation=True, 
+            max_length=512,
+            padding=True
+        )
+        
+        # Generate with optimized settings for medical text
         with torch.no_grad():
             outputs = model.generate(
                 **inputs,
-                max_new_tokens=512,
-                temperature=0.3,  # Lower temperature for medical precision
+                max_new_tokens=400,
+                temperature=0.2,  # Lower temperature for medical precision
+                top_p=0.9,
                 do_sample=True,
-                pad_token_id=processor.tokenizer.eos_token_id
+                repetition_penalty=1.1,
+                pad_token_id=tokenizer.eos_token_id,
+                eos_token_id=tokenizer.eos_token_id
             )
-
+        
         # Decode response
-        generated_text = processor.decode(outputs[0], skip_special_tokens=True)
-
-        # Extract only the generated part (remove the prompt)
-        soap_response = generated_text[len(prompt):].strip()
-
-        # Add header if requested
+        generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+        
+        # Extract only the SOAP note part
+        if "SOAP NOTE:" in generated_text:
+            soap_response = generated_text.split("SOAP NOTE:")[1].strip()
+        else:
+            soap_response = generated_text[len(prompt):].strip()
+        
+        # Clean up response
+        soap_response = clean_soap_response(soap_response)
+        
+        # Add professional header
         if include_timestamp:
             timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-            header = f"""📋 SOAP NOTE - Generated by Gemma 3N
+            header = f"""📋 SOAP NOTE - Generated by Gemma 3n
 🕐 Timestamp: {timestamp}
-🤖 Model: google/gemma-3n-e2b-it
-🔒 Processed locally on device: {device.upper()}
+🤖 Model: Unsloth-optimized Gemma 3n (4-bit quantized)
+🔒 Processed locally on device
+🏥 Medical Documentation Assistant
 
 {'='*60}
 """
             return header + soap_response
-
+        
         return soap_response
-
+        
     except Exception as e:
-        return f"❌ Error generating SOAP note: {str(e)}"
-
-print("✅ SOAP generation function ready!")
-
-"""## 📝 Interactive SOAP Note Generator
-### Enter medical notes below and generate professional SOAP documentation
-"""
-
-# Create interactive widgets
-print("🎨 Creating interactive interface...")
-
-# Text input area
-notes_input = widgets.Textarea(
-    value='',
-    placeholder='Enter unstructured doctor notes here...\n\nExample:\nPatient John Smith, 45yo male, came in complaining of chest pain for 2 days. Pain is sharp, 7/10 intensity, worse with movement. Vital signs: BP 140/90, HR 88, Temp 98.6F...',
-    description='Medical Notes:',
-    layout=widgets.Layout(width='100%', height='200px')
-)
-
-# File upload widget
-file_upload = widgets.FileUpload(
-    accept='.txt,.doc,.docx,.pdf',
-    multiple=False,
-    description='Or upload file:',
-    layout=widgets.Layout(width='300px')
-)
-
-# Generate button
-generate_btn = widgets.Button(
-    description='🤖 Generate SOAP Note',
-    button_style='primary',
-    layout=widgets.Layout(width='200px', height='40px')
-)
-
-# Output area
-output_area = widgets.HTML(
-    value='<p style="color: #666;">📋 Ready to generate SOAP notes! Enter medical notes above or upload a file.</p>',
-    layout=widgets.Layout(width='100%', height='400px', overflow='auto', border='1px solid #ddd', padding='10px')
-)
-
-# Example buttons
-example1_btn = widgets.Button(description='📋 Chest Pain Example', button_style='info', layout=widgets.Layout(width='180px'))
-example2_btn = widgets.Button(description='🩺 Diabetes Follow-up', button_style='info', layout=widgets.Layout(width='180px'))
-example3_btn = widgets.Button(description='👶 Pediatric Visit', button_style='info', layout=widgets.Layout(width='180px'))
-
-# Clear button
-clear_btn = widgets.Button(description='🗑️ Clear', button_style='warning', layout=widgets.Layout(width='100px'))
-
-print("✅ Interface widgets created!")
-
-# Example medical notes
-examples = {
-    'chest_pain': """Patient John Smith, 45yo male, came in complaining of chest pain for 2 days. Pain is sharp, 7/10 intensity, worse with movement. No radiation to arms. Vital signs: BP 140/90, HR 88, Temp 98.6F, RR 16, O2 sat 98%. Physical exam shows tenderness over left chest wall, no murmurs. EKG normal sinus rhythm. Chest X-ray clear. Diagnosed with costochondritis. Prescribed ibuprofen 600mg TID and advised rest. Follow up in 1 week if symptoms persist.""",
-
-    'diabetes': """Sarah Johnson, 62yo female with Type 2 diabetes, here for routine follow-up. Says blood sugars have been running high lately, 180-220 mg/dL. Taking metformin 1000mg BID. Diet has been poor due to holiday stress. Weight increased 5 lbs since last visit. BP 150/85, BMI 32. HbA1c 8.2% (was 7.1% 3 months ago). Feet exam normal, no neuropathy. Plan to increase metformin to 1000mg TID, refer to nutritionist, recheck labs in 3 months.""",
-
-    'pediatric': """Tommy Rodriguez, 8yo male, brought by mother for fever and cough x3 days. Fever up to 102F, productive cough with yellow sputum. Decreased appetite, no vomiting or diarrhea. Vital signs: Temp 101.2F, HR 110, RR 24, BP 95/60. Exam shows bilateral crackles in lower lobes, no wheeze. Throat clear. Diagnosed with bacterial pneumonia. Prescribed amoxicillin 500mg BID x10 days. Return if fever persists >48 hours on antibiotics."""
-}
-
-# Event handlers
-def on_generate_click(b):
-    with output_area:
-        output_area.value = '<p style="color: #007bff;">🔄 Processing with Gemma 3N... Please wait...</p>'
-
-        # Get text from input or uploaded file
-        text_to_process = notes_input.value
-
-        # Check if file was uploaded
-        if file_upload.value and len(file_upload.value) > 0:
-            try:
-                uploaded_file = list(file_upload.value.values())[0]
-                file_content = uploaded_file['content'].decode('utf-8')
-                text_to_process = file_content
-                notes_input.value = file_content  # Show in text area
-            except Exception as e:
-                output_area.value = f'<p style="color: #dc3545;">❌ Error reading file: {str(e)}</p>'
-                return
-
-        if not text_to_process.strip():
-            output_area.value = '<p style="color: #dc3545;">❌ Please enter medical notes or upload a file!</p>'
-            return
-
-        # Generate SOAP note
-        soap_note = generate_soap_note(text_to_process)
-
-        # Format output as HTML
-        formatted_output = f'<pre style="font-family: monospace; font-size: 12px; line-height: 1.4; white-space: pre-wrap;">{soap_note}</pre>'
-        output_area.value = formatted_output
-
-def on_example1_click(b):
-    notes_input.value = examples['chest_pain']
-    output_area.value = '<p style="color: #28a745;">✅ Chest pain example loaded! Click "Generate SOAP Note" to process.</p>'
-
-def on_example2_click(b):
-    notes_input.value = examples['diabetes']
-    output_area.value = '<p style="color: #28a745;">✅ Diabetes follow-up example loaded! Click "Generate SOAP Note" to process.</p>'
-
-def on_example3_click(b):
-    notes_input.value = examples['pediatric']
-    output_area.value = '<p style="color: #28a745;">✅ Pediatric example loaded! Click "Generate SOAP Note" to process.</p>'
-
-def on_clear_click(b):
-    notes_input.value = ''
-    file_upload.value = ()
-    output_area.value = '<p style="color: #666;">📋 Ready to generate SOAP notes! Enter medical notes above or upload a file.</p>'
-
-# Bind event handlers
-generate_btn.on_click(on_generate_click)
-example1_btn.on_click(on_example1_click)
-example2_btn.on_click(on_example2_click)
-example3_btn.on_click(on_example3_click)
-clear_btn.on_click(on_clear_click)
-
-print("✅ Event handlers configured!")
-
-# Define example medical notes first
-example_notes_1 = """
-Patient: John Smith, 45-year-old male
-Chief Complaint: Chest pain for 2 hours
-History: Patient reports sudden onset of sharp chest pain while at work. Pain is 7/10 intensity, located substernal, radiating to left arm. Associated with shortness of breath and diaphoresis. No previous cardiac history. Denies nausea or vomiting.
-Physical Exam: VS: BP 150/90, HR 110, RR 22, O2 Sat 96% on RA. Patient appears anxious and diaphoretic. Heart: Regular rhythm, no murmurs. Lungs: Clear bilaterally. Extremities: No edema.
-Assessment: Acute chest pain, rule out myocardial infarction
-Plan: EKG, cardiac enzymes, chest X-ray, aspirin 325mg, continuous cardiac monitoring
-"""
-
-example_notes_2 = """
-Patient: Sarah Johnson, 28-year-old female
-Chief Complaint: Severe headache and fever
-History: 3-day history of progressive headache, fever up to 101.5°F, photophobia, and neck stiffness. Patient reports this is the worst headache of her life. No recent travel or sick contacts. No rash noted.
-Physical Exam: VS: T 101.2°F, BP 130/80, HR 95, RR 18. Patient appears ill and photophobic. HEENT: Pupils equal and reactive. Neck: Stiff with positive Kernig's sign. Neurologic: Alert and oriented x3, no focal deficits.
-Assessment: Suspected meningitis
-Plan: Lumbar puncture, blood cultures, empiric antibiotics, supportive care
-"""
+        print(f"❌ Generation error: {e}")
+        return generate_template_soap(doctor_notes, include_timestamp)
+
+def clean_soap_response(response):
+    """Clean and format SOAP note response"""
+    
+    # Remove any incomplete sentences at the end
+    lines = response.split('\n')
+    cleaned_lines = []
+    
+    for line in lines:
+        line = line.strip()
+        if line:
+            # Ensure proper SOAP section headers
+            if line.upper().startswith(('SUBJECTIVE', 'OBJECTIVE', 'ASSESSMENT', 'PLAN')):
+                if not line.endswith(':'):
+                    line += ':'
+                cleaned_lines.append(f"\n{line}")
+            else:
+                cleaned_lines.append(line)
+    
+    return '\n'.join(cleaned_lines).strip()
+
+# Template-based SOAP generation (enhanced fallback)
+def generate_template_soap(doctor_notes, include_timestamp=True):
+    """Enhanced template-based SOAP note generation"""
+    
+    notes_lower = doctor_notes.lower()
+    lines = doctor_notes.split('\n')
+    
+    # Enhanced keyword extraction
+    subjective_info = extract_section_info(lines, [
+        'complains', 'reports', 'states', 'denies', 'pain', 'symptoms', 
+        'history', 'onset', 'duration', 'patient says', 'chief complaint'
+    ])
+    
+    objective_info = extract_section_info(lines, [
+        'vital signs', 'vs:', 'bp', 'hr', 'temp', 'examination', 'exam',
+        'physical', 'inspection', 'palpation', 'auscultation', 'laboratory'
+    ])
+    
+    assessment_info = extract_section_info(lines, [
+        'diagnosis', 'impression', 'assessment', 'likely', 'possible', 
+        'rule out', 'differential', 'icd', 'condition'
+    ])
+    
+    plan_info = extract_section_info(lines, [
+        'plan', 'treatment', 'medication', 'prescribe', 'follow', 'return',
+        'therapy', 'intervention', 'monitoring', 'referral'
+    ])
+    
+    # Build comprehensive SOAP note
+    soap_note = build_soap_sections(subjective_info, objective_info, assessment_info, plan_info)
+    
+    if include_timestamp:
+        timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        header = f"""📋 SOAP NOTE (Template-Enhanced)
+🕐 Timestamp: {timestamp}
+🔒 Processed locally - HIPAA compliant
+🏥 Scribbled Docs Medical Assistant
 
-example_notes_3 = """
-Patient: Robert Davis, 62-year-old male
-Chief Complaint: Shortness of breath and leg swelling
-History: 2-week history of progressive dyspnea on exertion, orthopnea, and bilateral lower extremity edema. Patient has history of hypertension and diabetes. Reports sleeping on 3 pillows due to breathing difficulty.
-Physical Exam: VS: BP 140/85, HR 88, RR 24, O2 Sat 92% on RA. Heart: S3 gallop present, JVD elevated. Lungs: Bilateral rales in lower fields. Extremities: 2+ pitting edema bilaterally.
-Assessment: Congestive heart failure exacerbation
-Plan: Chest X-ray, BNP, echocardiogram, furosemide, ACE inhibitor, daily weights
+{'='*60}
 """
+        return header + soap_note
+    
+    return soap_note
+
+def extract_section_info(lines, keywords):
+    """Extract relevant lines for each SOAP section"""
+    relevant_lines = []
+    for line in lines:
+        if any(keyword in line.lower() for keyword in keywords):
+            relevant_lines.append(line.strip())
+    return relevant_lines
+
+def build_soap_sections(subjective, objective, assessment, plan):
+    """Build formatted SOAP sections"""
+    
+    soap = "SUBJECTIVE:\n"
+    if subjective:
+        soap += '\n'.join(f"• {line}" for line in subjective[:5])  # Limit to 5 most relevant
+    else:
+        soap += "• Patient complaints and reported symptoms as documented"
+    
+    soap += "\n\nOBJECTIVE:\n"
+    if objective:
+        soap += '\n'.join(f"• {line}" for line in objective[:5])
+    else:
+        soap += "• Physical examination findings and clinical observations as documented"
+    
+    soap += "\n\nASSESSMENT:\n"
+    if assessment:
+        soap += '\n'.join(f"• {line}" for line in assessment[:3])
+    else:
+        soap += "• Clinical assessment based on presenting symptoms and examination findings"
+    
+    soap += "\n\nPLAN:\n"
+    if plan:
+        soap += '\n'.join(f"• {line}" for line in plan[:5])
+    else:
+        soap += "• Treatment plan and follow-up care as clinically indicated"
+    
+    return soap
 
-# Event handlers
-def on_generate_click(b):
-    try:
-        # Update the HTML widget directly
-        output_area.value = '<p style="color: #007bff;">🔄 Processing with Gemma 3N... Please wait...</p>'
-
-        # Get input text
-        input_text = notes_input.value.strip()
-
-        # Check if file was uploaded
-        if file_upload.value:
-            try:
-                # Process uploaded file
-                uploaded_file = list(file_upload.value.values())[0]
-                file_content = uploaded_file['content'].decode('utf-8')
-                input_text = file_content
-            except Exception as upload_error:
-                output_area.value = f'<p style="color: #ff6b6b;">❌ File upload error: {str(upload_error)}</p>'
-                return
-
-        if not input_text:
-            output_area.value = '<p style="color: #ff6b6b;">⚠️ Please enter medical notes or upload a file first!</p>'
-            return
-
-        # Check if generate_soap_note function exists
-        if 'generate_soap_note' not in globals():
-            output_area.value = '<p style="color: #ff6b6b;">❌ Error: generate_soap_note function not found. Please define it first.</p>'
-            return
-
-        # Generate SOAP note using Gemma
-        soap_note = generate_soap_note(input_text)
-
-        # Escape HTML in soap_note to prevent rendering issues
-        import html
-        escaped_soap_note = html.escape(soap_note)
-
-        # Display result
-        output_area.value = f'''
-        <div style="background: #f8f9fa; padding: 15px; border-radius: 8px; border-left: 4px solid #28a745;">
-            <h4 style="color: #28a745; margin-top: 0;">✅ Generated SOAP Note:</h4>
-            <pre style="white-space: pre-wrap; font-family: 'Courier New', monospace; background: white; padding: 15px; border-radius: 5px; border: 1px solid #ddd;">{escaped_soap_note}</pre>
-        </div>
-        '''
-
-    except Exception as e:
-        import traceback
-        error_details = traceback.format_exc()
-        output_area.value = f'''
-        <div style="color: #ff6b6b; background: #ffe6e6; padding: 15px; border-radius: 5px;">
-            <h4>❌ Error Details:</h4>
-            <p><strong>Error:</strong> {str(e)}</p>
-            <details>
-                <summary>Click for full traceback</summary>
-                <pre style="font-size: 12px; background: #fff; padding: 10px; border-radius: 3px; margin-top: 10px;">{error_details}</pre>
-            </details>
-        </div>
-        '''
-
-def on_clear_click(b):
+# OCR Functions (same as before but optimized)
+def initialize_ocr():
+    """Initialize OCR reader for handwritten notes"""
     try:
-        notes_input.value = ""
-        file_upload.value = ()
-        output_area.value = '<p>📋 Ready to generate SOAP notes! Enter medical notes above or upload a file.</p>'
+        # Initialize with English and medical text optimization
+        reader = easyocr.Reader(['en'], gpu=torch.cuda.is_available())
+        print("✅ EasyOCR initialized for handwritten medical notes")
+        return reader
     except Exception as e:
-        output_area.value = f'<p style="color: #ff6b6b;">❌ Clear error: {str(e)}</p>'
-
-def on_example_click(example_text):
-    def handler(b):
-        try:
-            notes_input.value = example_text
-            output_area.value = '<p style="color: #28a745;">📋 Example loaded! Click "Generate SOAP Note" to process.</p>'
-        except Exception as e:
-            output_area.value = f'<p style="color: #ff6b6b;">❌ Example load error: {str(e)}</p>'
-    return handler
-
-# Connect event handlers to buttons
-try:
-    generate_btn.on_click(on_generate_click)
-    clear_btn.on_click(on_clear_click)
-    example1_btn.on_click(on_example_click(example_notes_1))
-    example2_btn.on_click(on_example_click(example_notes_2))
-    example3_btn.on_click(on_example_click(example_notes_3))
-
-    print("✅ Event handlers connected successfully!")
-    print("📋 Example notes loaded:")
-    print("   - Example 1: Chest pain case")
-    print("   - Example 2: Suspected meningitis")
-    print("   - Example 3: Heart failure")
-
-except Exception as e:
-    print(f"❌ Error connecting event handlers: {str(e)}")
-    import traceback
-    traceback.print_exc()
-
-"""## 🌐 Alternative: Gradio Web Interface
-### Run this cell for a shareable web interface
-"""
-
-# Install required packages for image processing and OCR
-
-import gradio as gr
-import torch
-from PIL import Image
-import pytesseract
-import cv2
-import numpy as np
-import easyocr
-import io
-
-# First, make sure you have the examples dictionary defined
-examples = {
-    'chest_pain': """Patient: John Smith, 45-year-old male
-Chief Complaint: Chest pain for 2 hours
-History: Patient reports sudden onset of sharp chest pain while at work. Pain is 7/10 intensity, located substernal, radiating to left arm. Associated with shortness of breath and diaphoresis. No previous cardiac history. Denies nausea or vomiting.
-Physical Exam: VS: BP 150/90, HR 110, RR 22, O2 Sat 96% on RA. Patient appears anxious and diaphoretic. Heart: Regular rhythm, no murmurs. Lungs: Clear bilaterally. Extremities: No edema.
-Assessment: Acute chest pain, rule out myocardial infarction
-Plan: EKG, cardiac enzymes, chest X-ray, aspirin 325mg, continuous cardiac monitoring""",
-
-    'diabetes': """Patient: Maria Garcia, 52-year-old female
-Chief Complaint: Increased thirst and frequent urination for 3 weeks
-History: Patient reports polyuria, polydipsia, and unintentional weight loss of 10 lbs over past month. Family history of diabetes. Denies fever, abdominal pain, or vision changes.
-Physical Exam: VS: BP 140/85, HR 88, RR 16, BMI 28. Patient appears well but slightly dehydrated. HEENT: Dry mucous membranes. Cardiovascular: Regular rate and rhythm. Extremities: No diabetic foot changes noted.
-Assessment: New onset diabetes mellitus, likely Type 2
-Plan: HbA1c, fasting glucose, comprehensive metabolic panel, diabetic education, metformin initiation""",
-
-    'pediatric': """Patient: Emma Thompson, 8-year-old female
-Chief Complaint: Fever and sore throat for 2 days
-History: Mother reports fever up to 102°F, sore throat, difficulty swallowing, and decreased appetite. No cough or runny nose. Several classmates have been sick with similar symptoms.
-Physical Exam: VS: T 101.8°F, HR 110, RR 20, O2 Sat 99%. Patient appears mildly ill but alert. HEENT: Throat erythematous with tonsillar exudate, anterior cervical lymphadenopathy. Heart and lungs: Normal.
-Assessment: Streptococcal pharyngitis (probable)
-Plan: Rapid strep test, throat culture, amoxicillin if positive, supportive care, return if worsening"""
-}
-
-# Initialize EasyOCR reader (better for handwritten text)
-try:
-    ocr_reader = easyocr.Reader(['en'])
-    print("✅ EasyOCR initialized successfully")
-except:
-    ocr_reader = None
-    print("⚠️ EasyOCR not available, using Tesseract only")
-
-def preprocess_image_for_ocr(image):
-    """
-    Preprocess image to improve OCR accuracy
-    """
-    # Convert PIL Image to numpy array
-    img_array = np.array(image)
-
-    # Convert to grayscale if needed
-    if len(img_array.shape) == 3:
-        gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)
-    else:
-        gray = img_array
-
-    # Apply image preprocessing for better OCR
-    # 1. Resize image if too small
-    height, width = gray.shape
-    if height < 300 or width < 300:
-        scale_factor = max(300/height, 300/width)
-        new_width = int(width * scale_factor)
-        new_height = int(height * scale_factor)
-        gray = cv2.resize(gray, (new_width, new_height), interpolation=cv2.INTER_CUBIC)
-
-    # 2. Noise removal
-    denoised = cv2.medianBlur(gray, 3)
+        print(f"⚠️ EasyOCR initialization failed: {e}")
+        return None
 
-    # 3. Contrast enhancement
-    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
-    enhanced = clahe.apply(denoised)
-
-    # 4. Thresholding
-    _, thresh = cv2.threshold(enhanced, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
-
-    return thresh
-
-def extract_text_from_image(image):
-    """
-    Extract text from image using multiple OCR methods
-    """
+def extract_text_from_image(image, ocr_reader=None):
+    """Enhanced OCR for medical handwriting"""
     if image is None:
         return "❌ No image provided"
-
+    
     try:
-        # Preprocess image
-        processed_img = preprocess_image_for_ocr(image)
-
-        # Method 1: Try EasyOCR (better for handwritten text)
+        # Preprocess specifically for medical handwriting
+        processed_img = preprocess_medical_image(image)
+        
+        extracted_text = ""
+        
+        # Try EasyOCR (better for handwritten text)
         if ocr_reader is not None:
             try:
-                # Convert back to PIL Image for EasyOCR
-                pil_img = Image.fromarray(processed_img)
-                results = ocr_reader.readtext(np.array(pil_img))
-
-                # Extract text from EasyOCR results
-                easyocr_text = ' '.join([result[1] for result in results])
-
-                if len(easyocr_text.strip()) > 20:  # If we got good results
-                    return clean_extracted_text(easyocr_text)
-
+                results = ocr_reader.readtext(processed_img, detail=0, paragraph=True)
+                if results:
+                    extracted_text = ' '.join(results)
+                    if len(extracted_text.strip()) > 10:
+                        return clean_medical_text(extracted_text)
             except Exception as e:
                 print(f"EasyOCR failed: {e}")
-
-        # Method 2: Tesseract OCR (fallback)
+        
+        # Fallback to Tesseract with medical optimization
         try:
-            # Configure Tesseract for medical text
-            custom_config = r'--oem 3 --psm 6'
+            import pytesseract
+            
+            # Medical-optimized Tesseract config
+            custom_config = r'--oem 3 --psm 6 -c tessedit_char_whitelist=ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789.,;:()[]{}/-+= '
+            
             tesseract_text = pytesseract.image_to_string(processed_img, config=custom_config)
-
-            if len(tesseract_text.strip()) > 10:
-                return clean_extracted_text(tesseract_text)
-
+            
+            if len(tesseract_text.strip()) > 5:
+                return clean_medical_text(tesseract_text)
+                
         except Exception as e:
             print(f"Tesseract failed: {e}")
-
-        return "❌ Could not extract text from image. Please try a clearer image or enter text manually."
-
+        
+        return "❌ Could not extract text from image. Please ensure the image is clear and try again."
+        
     except Exception as e:
         return f"❌ Error processing image: {str(e)}"
 
-def clean_extracted_text(text):
-    """
-    Clean up extracted text
-    """
+def preprocess_medical_image(image):
+    """Optimized preprocessing for medical handwriting"""
+    try:
+        img_array = np.array(image)
+        
+        # Convert to grayscale
+        if len(img_array.shape) == 3:
+            gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)
+        else:
+            gray = img_array
+        
+        # Resize for optimal OCR (medical notes are often small)
+        height, width = gray.shape
+        if height < 400 or width < 400:
+            scale_factor = max(400/height, 400/width)
+            new_width = int(width * scale_factor)
+            new_height = int(height * scale_factor)
+            gray = cv2.resize(gray, (new_width, new_height), interpolation=cv2.INTER_CUBIC)
+        
+        # Advanced preprocessing for handwritten medical text
+        # 1. Noise reduction
+        denoised = cv2.fastNlMeansDenoising(gray)
+        
+        # 2. Contrast enhancement specifically for handwriting
+        clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8,8))
+        enhanced = clahe.apply(denoised)
+        
+        # 3. Morphological operations to clean up text
+        kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1,1))
+        cleaned = cv2.morphologyEx(enhanced, cv2.MORPH_CLOSE, kernel)
+        
+        # 4. Adaptive thresholding (better for varying lighting)
+        thresh = cv2.adaptiveThreshold(
+            cleaned, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2
+        )
+        
+        return thresh
+        
+    except Exception as e:
+        print(f"❌ Image preprocessing error: {e}")
+        return np.array(image)
+
+def clean_medical_text(text):
+    """Clean extracted text with medical context awareness"""
     # Remove excessive whitespace and empty lines
     lines = [line.strip() for line in text.split('\n') if line.strip()]
-    cleaned_text = '\n'.join(lines)
-
-    # Remove special characters that might interfere
-    cleaned_text = cleaned_text.replace('|', '').replace('_', ' ')
-
-    return cleaned_text.strip()
-
-def gradio_generate_soap(medical_notes, uploaded_image):
-    """
-    Modified Gradio interface function for SOAP generation from images
-    """
+    
+    # Medical text cleaning
+    cleaned_lines = []
+    for line in lines:
+        # Remove obvious OCR artifacts
+        line = line.replace('|', 'l').replace('_', ' ').replace('~', '-')
+        
+        # Fix common medical abbreviations that OCR might misread
+        medical_corrections = {
+            'BP': 'BP', 'HR': 'HR', 'RR': 'RR', 'O2': 'O2',
+            'mg': 'mg', 'ml': 'ml', 'cc': 'cc', 'cm': 'cm'
+        }
+        
+        for wrong, correct in medical_corrections.items():
+            line = line.replace(wrong.lower(), correct)
+        
+        if len(line) > 1:  # Filter out single characters
+            cleaned_lines.append(line)
+    
+    return '\n'.join(cleaned_lines)
+
+# Enhanced Gradio Interface
+def gradio_generate_soap(medical_notes, uploaded_image, model_data):
+    """Main Gradio interface function"""
+    model, tokenizer = model_data if model_data else (None, None)
+    ocr_reader = getattr(gradio_generate_soap, 'ocr_reader', None)
+    
     text_to_process = medical_notes.strip() if medical_notes else ""
-
-    # If image is uploaded, extract text using OCR
+    
+    # Process uploaded image with enhanced OCR
     if uploaded_image is not None:
         try:
-            print("🔍 Extracting text from uploaded image...")
-            extracted_text = extract_text_from_image(uploaded_image)
-
-            # Check if OCR was successful
-            if extracted_text.startswith("❌"):
-                return extracted_text
-
-            # Use extracted text if manual text is empty or append to manual text
-            if not text_to_process:
-                text_to_process = extracted_text
+            print("🔍 Extracting text from medical image...")
+            extracted_text = extract_text_from_image(uploaded_image, ocr_reader)
+            
+            if not extracted_text.startswith("❌"):
+                if not text_to_process:
+                    text_to_process = f"--- Extracted from uploaded image ---\n{extracted_text}"
+                else:
+                    text_to_process = f"{text_to_process}\n\n--- Additional text from image ---\n{extracted_text}"
             else:
-                text_to_process = f"{text_to_process}\n\n--- Extracted from image ---\n{extracted_text}"
-
+                return extracted_text
+                
         except Exception as e:
             return f"❌ Error processing image: {str(e)}"
-
+    
     if not text_to_process:
-        return "❌ Please enter medical notes manually or upload a PNG/JPG image with medical text"
-
-    # Check if generate_soap_note function exists
-    if 'generate_soap_note' not in globals():
-        return "❌ Error: generate_soap_note function not found. Please define it first."
-
+        return "❌ Please enter medical notes manually or upload an image with medical text"
+    
+    # Generate SOAP note using Gemma 3n
     try:
-        return generate_soap_note(text_to_process)
+        return generate_soap_note_gemma(text_to_process, model, tokenizer)
     except Exception as e:
         return f"❌ Error generating SOAP note: {str(e)}"
 
-# Create example images (you can replace these with actual medical note images)
-def create_example_image(text, filename):
-    """
-    Create example images from text (for demonstration)
-    """
-    from PIL import Image, ImageDraw, ImageFont
-
-    # Create a white image
-    img = Image.new('RGB', (800, 600), color='white')
-    draw = ImageDraw.Draw(img)
-
-    try:
-        # Try to use a default font
-        font = ImageFont.load_default()
-    except:
-        font = None
-
-    # Add text to image
-    lines = text.split('\n')
-    y_offset = 20
-    for line in lines[:15]:  # Limit to first 15 lines
-        draw.text((20, y_offset), line, fill='black', font=font)
-        y_offset += 25
-
-    return img
-
-# Create Gradio interface
-gradio_interface = gr.Interface(
-    fn=gradio_generate_soap,
-    inputs=[
-        gr.Textbox(
-            lines=6,
-            placeholder="Enter medical notes manually (optional)...\n\nOr upload an image below and text will be extracted automatically.",
-            label="📝 Medical Notes (Manual Entry)"
-        ),
-        gr.Image(
-            type="pil",
-            label="📷 Upload Medical Image (PNG/JPG only)",
-            sources=["upload", "webcam"],  # FIXED: Changed "camera" to "webcam"
-            image_mode="RGB"
-        )
-    ],
-    outputs=[
-        gr.Textbox(
-            lines=15,
-            label="📋 Generated SOAP Note",
-            show_copy_button=True
-        )
-    ],
-    title="🏥 Medical Image SOAP Note Generator",
-    description="""
-    Transform medical images (PNG/JPG) into professional SOAP documentation using OCR + Gemma 3N model.
-
-    📸 **How to use:**
-    1. Upload a PNG or JPG image of medical notes (typed or handwritten)
-    2. Or enter text manually in the text box above
-    3. The system will extract text from images using OCR
-    4. Generate structured SOAP notes automatically
-
-    💡 **Tips for better OCR results:**
-    - Use clear, high-resolution images
-    - Ensure good lighting and contrast
-    - Keep text horizontal (not tilted)
-    - Handwritten text works best when clearly written
-    """,
-    examples=[
-        [examples['chest_pain'], None],
-        [examples['diabetes'], None],
-        [examples['pediatric'], None]
-    ],
-    theme=gr.themes.Soft(),
-    flagging_mode="never"
-)
-
-# Launch Gradio interface with flexible port selection
-print("🚀 Launching Medical Image SOAP Generator...")
+# Example medical notes for testing
+medical_examples = {
+    'chest_pain': """Patient: John Smith, 45yo M
+CC: Chest pain x 2 hours
+HPI: Sudden onset sharp substernal chest pain 7/10, radiating to L arm. Associated SOB, diaphoresis. No N/V.
+PMH: HTN, no CAD
+VS: BP 150/90, HR 110, RR 22, O2 96% RA
+PE: Anxious, diaphoretic. RRR, no murmur. CTAB. No edema.
+A: Acute chest pain, r/o MI
+P: EKG, troponins, CXR, ASA 325mg, monitor""",
+
+    'diabetes': """Patient: Maria Garcia, 52yo F
+CC: Increased thirst, urination x 3 weeks  
+HPI: Polyuria, polydipsia, 10lb weight loss. FH DM. No fever, abd pain.
+VS: BP 140/85, HR 88, BMI 28
+PE: Mild dehydration, dry MM. RRR. No diabetic foot changes.
+Labs: Random glucose 280, HbA1c pending
+A: New onset DM Type 2
+P: HbA1c, CMP, diabetic education, metformin, f/u 2 weeks""",
+
+    'pediatric': """Patient: Emma Thompson, 8yo F
+CC: Fever, sore throat x 2 days
+HPI: Fever 102F, sore throat, odynophagia, decreased appetite. No cough/rhinorrhea.
+VS: T 101.8F, HR 110, RR 20, O2 99%
+PE: Alert, mildly ill. Throat erythematous w/ tonsillar exudate. Anterior cervical LAD.
+A: Strep pharyngitis (probable)
+P: Rapid strep, throat culture, amoxicillin if +, supportive care, RTC PRN"""
+}
 
-try:
-    # Try different ports if 7860 is busy
-    for port in [7860, 7861, 7862, 7863, 7864]:
-        try:
-            gradio_interface.launch(
-                share=True,  # Creates a public shareable link
-                server_port=port,
-                show_error=True,
-                quiet=False
+# Initialize everything
+def initialize_app():
+    """Initialize the complete application"""
+    print("🚀 Initializing Scribbled Docs SOAP Generator...")
+    
+    # Setup device
+    device = setup_device()
+    
+    # Load model
+    model, tokenizer = load_unsloth_gemma_model(device)
+    
+    # Initialize OCR
+    ocr_reader = initialize_ocr()
+    gradio_generate_soap.ocr_reader = ocr_reader
+    
+    return model, tokenizer
+
+# Create the main Gradio interface
+def create_interface(model, tokenizer):
+    """Create the main Gradio interface"""
+    
+    interface = gr.Interface(
+        fn=lambda notes, image: gradio_generate_soap(notes, image, (model, tokenizer)),
+        inputs=[
+            gr.Textbox(
+                lines=8,
+                placeholder="Enter medical notes here...\n\nExample:\nPatient: John Doe, 45yo M\nCC: Chest pain\nVS: BP 140/90, HR 88\n...",
+                label="📝 Medical Notes (Manual Entry)",
+                info="Enter unstructured medical notes or upload an image below"
+            ),
+            gr.Image(
+                type="pil",
+                label="📷 Upload Medical Image (Handwritten/Typed Notes)",
+                sources=["upload", "webcam"],
+                info="Upload PNG/JPG images of medical notes - handwritten or typed"
             )
-            print(f"✅ Interface launched successfully on port {port}")
-            break
-        except OSError as port_error:
-            print(f"⚠️ Port {port} is busy, trying next port...")
-            continue
-    else:
-        # If all ports are busy, let Gradio choose automatically
-        print("🔄 All preferred ports busy, letting Gradio choose automatically...")
-        gradio_interface.launch(
-            share=True,
+        ],
+        outputs=[
+            gr.Textbox(
+                lines=20,
+                label="📋 Generated SOAP Note",
+                show_copy_button=True,
+                info="Professional SOAP note generated from your input"
+            )
+        ],
+        title="🏥 Scribbled Docs - Medical SOAP Note Generator",
+        description="""
+        **Transform medical notes into professional SOAP documentation using Gemma 3n AI**
+        
+        🔒 **100% Offline & HIPAA Compliant** - All processing happens locally on your device
+        🤖 **Powered by Unsloth-optimized Gemma 3n** - 4-bit quantized for efficiency
+        📝 **Supports handwritten & typed notes** - Advanced OCR for medical handwriting
+        
+        **Instructions:**
+        1. Enter medical notes manually OR upload an image
+        2. Click Submit to generate a structured SOAP note
+        3. Copy the result for use in your medical records
+        
+        **Perfect for:** Emergency medicine, family practice, internal medicine, pediatrics
+        """,
+        examples=[
+            [medical_examples['chest_pain'], None],
+            [medical_examples['diabetes'], None], 
+            [medical_examples['pediatric'], None]
+        ],
+        theme=gr.themes.Soft(
+            primary_hue="blue",
+            secondary_hue="green"
+        ),
+        allow_flagging="never",
+        analytics_enabled=False
+    )
+    
+    return interface
+
+# Main execution
+if __name__ == "__main__":
+    try:
+        # Initialize app
+        model, tokenizer = initialize_app()
+        
+        # Create and launch interface
+        interface = create_interface(model, tokenizer)
+        
+        print("\n🎯 Scribbled Docs SOAP Generator Ready!")
+        print("📱 Features:")
+        print("   ✅ Offline processing (HIPAA compliant)")
+        print("   ✅ Unsloth-optimized Gemma 3n model")
+        print("   ✅ Handwritten note OCR")
+        print("   ✅ Professional SOAP formatting")
+        print("   ✅ Medical terminology aware")
+        
+        # Launch interface
+        interface.launch(
+            share=True,  # Creates public link
+            server_port=7860,
             show_error=True,
             quiet=False
         )
-
-except Exception as e:
-    print(f"❌ Error launching Gradio interface: {str(e)}")
-    print("💡 Alternative: Try running without share=True:")
-    print("gradio_interface.launch(show_error=True)")
-
-print("🎯 Medical Image SOAP Generator ready!")
-print("📸 Upload PNG/JPG images of medical notes for automatic text extraction and SOAP generation")
-
-"""## 📊 Usage Statistics & Model Info"""
-
-# Display model and system information
-import psutil
-import GPUtil
-
-def show_system_info():
-    print("🔧 SYSTEM INFORMATION")
-    print("="*50)
-    print(f"🖥️  Device: {device.upper()}")
-    print(f"🧠 CPU Usage: {psutil.cpu_percent(interval=1):.1f}%")
-    print(f"💾 RAM Usage: {psutil.virtual_memory().percent:.1f}%")
-
-    if torch.cuda.is_available():
-        try:
-            gpus = GPUtil.getGPUs()
-            if gpus:
-                gpu = gpus[0]
-                print(f"🚀 GPU: {gpu.name}")
-                print(f"📊 GPU Usage: {gpu.load*100:.1f}%")
-                print(f"🔥 GPU Memory: {gpu.memoryUsed}/{gpu.memoryTotal} MB ({gpu.memoryPercent:.1f}%)")
-                print(f"🌡️  GPU Temp: {gpu.temperature}°C")
-        except:
-            print(f"🚀 GPU Memory: {torch.cuda.memory_allocated()/1e9:.1f}GB / {torch.cuda.memory_reserved()/1e9:.1f}GB")
-
-    print("\n🤖 MODEL INFORMATION")
-    print("="*50)
-    print(f"📡 Model ID: {model_id}")
-    print(f"🎯 Model Type: Multimodal (Text, Image, Audio)")
-    print(f"📊 Model Size: ~2.9GB")
-    print(f"🔢 Parameters: ~2.9B")
-    print(f"🌍 Languages: 140 text + 35 multimodal")
-    print(f"💽 Precision: {model.dtype}")
-
-    print("\n✅ Ready for SOAP note generation!")
-
-show_system_info()
-
-"""---
-## 📋 SOAP Note Format Reference
-
-**S - SUBJECTIVE**: Patient's reported symptoms and history  
-**O - OBJECTIVE**: Observable clinical findings  
-**A - ASSESSMENT**: Clinical diagnosis/impression  
-**P - PLAN**: Treatment and follow-up plan  
-
----
-*🤖 Powered by Google's Gemma 3N Model | 🔒 All processing performed locally*
-"""
-
+        
+    except Exception as e:
+        print(f"❌ Error launching application: {e}")
+        print("💡 Make sure you have installed: pip install unsloth gradio easyocr opencv-python")