app.py

import os
import ast
import spaces
import gradio as gr

from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import login as hf_login

import xgrammar as xgr
from pydantic import BaseModel

hf_login(token=os.getenv("HF_TOKEN"))

model_name = "gregorlied/Llama-3.2-1B-Instruct-Medical-Report-Summarization-FP32"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    attn_implementation='eager',
    trust_remote_code=True,
)

class Person(BaseModel):
  life_style: str
  family_history: str
  social_history: str
  medical_surgical_history: str
  signs_symptoms: str
  comorbidities: str
  diagnostic_techniques_procedures: str
  diagnosis: str
  laboratory_values: str
  pathology: str
  pharmacological_therapy: str
  interventional_therapy: str
  patient_outcome_assessment: str
  age: str
  gender: str
    
config = AutoConfig.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

tokenizer_info = xgr.TokenizerInfo.from_huggingface(
    tokenizer, vocab_size=len(tokenizer)
)

grammar_compiler = xgr.GrammarCompiler(tokenizer_info)
compiled_grammar = grammar_compiler.compile_json_schema(Person)

default_value = "A 57-year-old male presented with fever (38.9°C), chest pain, cough, and progressive dyspnea. The patient exhibited tachypnea (34 breaths/min) and tachycardia (134 bpm). Auscultation revealed decreased breath sounds in both lung bases, with crackles on the left. A chest X-ray revealed bilateral pleural opacities and enlargement of the cardiac silhouette ( A). Echocardiography showed moderate pericardial effusion affecting the entire cardiac silhouette. Pericardiocentesis yielded 250 mL of exudative fluid. A CT scan of the chest showed pneumonia in the left lower lobe, bilateral pleural effusion, and moderate pericardial effusion ( B). Thoracentesis was performed and yielded 1,050 mL of exudative fluid. Laboratory tests yielded the following data: white blood cell count, 11.78 × 109 cells/L (84.3% neutrophils, 4.3% lymphocytes, and 9.1% monocytes); platelet count, 512 × 109/L; serum C-reactive protein, 31.27 mg/dL; serum creatinine, 0.94 mg/dL; serum sodium, 133 mEq/L; and serum potassium, 3.72 mEq/L. Examination of the pleural fluid showed a pH of 7.16, a glucose level of 4.5 mg/dL, proteins at 49.1 g/L, and an LDH content of 1,385 U/L. A urinary pneumococcal antigen test was positive. Pleural fluid culture was positive for S. pneumoniae. The patient was treated for four weeks with amoxicillin-clavulanate (2.2 g/8 h, i.v.) plus levofloxacin (500 mg twice a day), together with a nonsteroidal anti-inflammatory drug (ibuprofen, 800 mg/day), after which there was nearly complete resolution of the alterations seen on the chest X-ray and CT scan."

prompt = """You are a text extraction system for clinical reports.
Please extract relevant clinical information from the report.

### Instructions

- Use the JSON Schema given below.
- Return only a valid JSON object – no markdown, no comments.
- If no relevant facts are given for a field, set its value to "N/A".
- If multile relevant facts are given for a field, separate them with "; ".

### JSON Schema

{
  'life_style': '',
  'family_history': '',
  'social_history': '',
  'medical_surgical_history': '',
  'signs_symptoms': '',
  'comorbidities': '',
  'diagnostic_techniques_procedures': '',
  'diagnosis': '',
  'laboratory_values': '',
  'pathology': '',
  'pharmacological_therapy': '',
  'interventional_therapy': '',
  'patient_outcome_assessment': '',
  'age': '',
  'gender': '',
}

### Clinical Report
"""

def generate_html_tables(data, selected_fields):
    key_label_map = {
        'age': 'Age',
        'gender': 'Gender',
        'life_style': 'Lifestyle',
        'social_history': 'Social Background',
        'medical_surgical_history': 'Personal',
        'family_history': 'Family Members',
        'signs_symptoms': 'Symptoms',
        'comorbidities': 'Comorbid Conditions',
        'diagnostic_techniques_procedures': 'Diagnostic Procedures',
        'laboratory_values': 'Laboratory Results',
        'pathology': 'Pathology Report',
        'diagnosis': 'Diagnosis',
        'interventional_therapy': 'Interventional Therapy',
        'pharmacological_therapy': 'Pharmacological Therapy',
        'patient_outcome_assessment': 'Patient Outcome',
    }
    
    label_key_map = {v: k for k, v in key_label_map.items()}

    categories = {
        "Personal Information": ["Age", "Gender", "Lifestyle", "Social Background"],
        "Medical History": ["Personal", "Family Members"],
        "Clinical Presentation": ["Symptoms", "Comorbid Conditions"],
        "Medical Assessment": ["Diagnostic Procedures", "Laboratory Results", "Pathology Report"],
        "Diagnosis": ["Diagnosis"],
        "Treatment": ["Interventional Therapy", "Pharmacological Therapy"],
        "Patient Outcome": ["Patient Outcome"],
    }

    def format_bullets(value):
        items = [item.strip() for item in value.split(";") if item.strip()]
        if not items:
            return "<i>Not Available</i>"
        if len(items) == 1:
            return items[0]
        return "<ul style='margin: 0; padding-left: 1em'>" + "".join(f"<li>{item}</li>" for item in items) + "</ul>"

    table_style = (
        "width: 100%;"
        "height: 100%;"
        "table-layout: fixed;"
    )

    th_td_style = (
        "padding: 8px;"
        "border: 1px solid #ccc;"
        "vertical-align: top;"
        "text-align: left;"
    )

    html_tables = []
    for section, labels in categories.items():
        section_fields = [label for label in labels if label in selected_fields]
        if section_fields:
            table_html = f"<h3 style='margin-bottom: 0.5em;'>{section}</h3>"
            table_html += f"<table style='{table_style}'>"
            table_html += f"<tr><th style='height: 30px; {th_td_style}; width: 150px;'>Field</th><th style='height: 30px; {th_td_style};'>Details</th></tr>"
            for label in section_fields:
                key = label_key_map[label]
                value = data.get(key, "N/A")
                details = "<i>Not Available</i>" if value == "N/A" else format_bullets(value)
                table_html += f"<tr><td style='{th_td_style}; width: 150px;'><b>{label}</b></td><td style='{th_td_style}'>{details}</td></tr>"
            table_html += "</table>"
            html_tables.append(table_html)

    i = 0
    grouped_html = ""
    while i < len(html_tables):
        num_per_row = 2 if i < 4 else 3
        row_tables = html_tables[i:i+num_per_row]
        grouped_html += (
            "<div style='display: flex; gap: 1em; margin-bottom: 2em;'>"
        )
        for table in row_tables:
            grouped_html += (
                "<div style='display: flex; flex-direction: column;'>"
                f"{table}"
                "</div>"
            )
        grouped_html += "</div>"
        i += num_per_row

    return f"<div style='font-family: sans-serif;'>{grouped_html}</div>"

@spaces.GPU(duration=60)
def summarize(
    text, 
    personal_info, 
    medical_history, 
    clinical_presentation, 
    medical_assessment, 
    diagnosis, 
    treatment, 
    patient_outcome,
):
    
    if not text.strip():
        return "Please enter some text to summarize."

    if text == default_value: 
        response = ['{"life_style": "N/A", "family_history": "N/A", "social_history": "N/A", "medical_surgical_history": "N/A", "signs_symptoms": "Fever; Chest pain; Cough; Progressive dyspnea; Tachypnea; Tachycardia; Decreased breath sounds in both lung bases; Crackles on the left", "comorbidities": "N/A", "diagnostic_techniques_procedures": "Chest X-ray; Echocardiography; Thoracentesis; Laboratory tests; Pleural fluid analysis; Urinary pneumococcal antigen test; Pleural fluid culture", "diagnosis": "Pneumonia; Pericardial effusion; S. pneumoniae infection", "laboratory_values": "White blood cell count: 11.78 \\u00d7 10^9 cells/L (84.3% neutrophils, 4.3% lymphocytes, 9.1% monocytes); Platelet count: 512 \\u00d7 10^9/L; Serum C-reactive protein: 31.27 mg/dL; Serum creatinine: 0.94 mg/dL; Serum sodium: 133 mEq/L; Serum potassium: 3.72 mEq/L; Pleural fluid pH: 7.16; Pleural fluid glucose: 4.5 mg/dL; Pleural fluid proteins: 49.1 g/L; Pleural fluid LDH: 1,385 U/L", "pathology": "N/A", "pharmacological_therapy": "Amoxicillin-clavulanate (2.2 g/8 h, i.v.); Levofloxacin (500 mg twice a day); Ibuprofen (800 mg/day)", "interventional_therapy": "Pericardiocentesis; Thoracentesis", "patient_outcome_assessment": "Nearly complete resolution of alterations on chest X-ray and CT scan", "age": "57 year", "gender": "Male"}']
    else:
        messages = [
            {"role": "system", "content": prompt.strip()},
            {"role": "user", "content": text.strip()},
        ]

        text = tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True,
            enable_thinking=False, # only relevant for qwen
        )

        # We cannot reset here because __call__ is not invoked when stop token is sampled.
        # Therefore, each `generate()` call needs to instantiate an LogitsProcessor.
        xgr_logits_processor = xgr.contrib.hf.LogitsProcessor(compiled_grammar)
    
        model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
        generated_ids = model.generate(
            input_ids=model_inputs["input_ids"],
            attention_mask = model_inputs["attention_mask"],
            max_new_tokens=2048,
            logits_processor=[xgr_logits_processor]
        )
        generated_ids = [
            output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
        ]
    
        response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)

    try:
        data = ast.literal_eval(response[0])
    except:
        data = {
            'life_style': 'N/A',
            'family_history': 'N/A',
            'social_history': 'N/A',
            'medical_surgical_history': 'N/A',
            'signs_symptoms': 'N/A',
            'comorbidities': 'N/A',
            'diagnostic_techniques_procedures': 'N/A',
            'diagnosis': 'N/A',
            'laboratory_values': 'N/A',
            'pathology': '',
            'pharmacological_therapy': 'N/A',
            'interventional_therapy': 'N/A',
            'patient_outcome_assessment': 'N/A',
            'age': 'N/A',
            'gender': 'N/A',
        }
        
    selected_fields = []
    selected_fields += personal_info
    selected_fields += medical_history
    selected_fields += clinical_presentation
    selected_fields += medical_assessment
    selected_fields += diagnosis
    selected_fields += treatment
    selected_fields += patient_outcome
    
    return generate_html_tables(data, selected_fields)

with gr.Blocks() as demo:

    # need to be combined with `hf_oauth: true` in README.md
    # button = gr.LoginButton("Sign in")

    with gr.Column():
        gr.HTML("""
<div align="center">
  <img src="https://huggingface.co/spaces/gregorlied/medical-text-summarization/resolve/main/assets/LlamaMD-logo.png" alt="LlamaMD Logo" width="120" style="margin-bottom: 10px;">
  <h2><strong>LlamaMD</strong></h2>
  <p><em>Structured Information Extraction from Clinical Reports</em></p>
</div>
""")

    with gr.Tabs():
        with gr.Tab("LLamaMD"):
            with gr.Row():
                input_text = gr.Textbox(
                    label="Clinical Report", 
                    autoscroll=False,
                    lines=15, 
                    max_lines=15, 
                    placeholder="Paste your clinical report here...",
                    value=default_value,
                )
        
            with gr.Accordion("Advanced Settings", open=False):
                with gr.Row():
                    with gr.Column():
                        personal_info = gr.CheckboxGroup(
                            label="Personal Information", 
                            choices=[
                                "Age",
                                "Gender",
                                "Lifestyle", 
                                "Social Background",
                            ],
                            value=[
                                "Age",
                                "Gender",
                                "Lifestyle", 
                                "Social Background",
                            ],
                        )
                    
                    with gr.Column():
                        medical_history = gr.CheckboxGroup(
                            label="Medical History", 
                            choices=[ 
                                "Personal",
                                "Family Members",
                            ],
                            value=[
                                "Personal",
                                "Family Members",
                            ],
                        )
        
                with gr.Row():
                    with gr.Column():
                        clinical_presentation = gr.CheckboxGroup(
                            label="Clinical Presentation", 
                            choices=[
                                "Symptoms",
                                "Comorbid Conditions",
                            ],
                            value=[
                                "Symptoms",
                                "Comorbid Conditions",
                            ],
                        )
                    with gr.Column():
                        medical_assessment = gr.CheckboxGroup(
                            label="Medical Assessment", 
                            choices=[
                                "Diagnostic Procedures",
                                "Laboratory Results",
                                "Pathology Report",       
                            ],
                            value=[
                                "Diagnostic Procedures",
                                "Laboratory Results",
                                "Pathology Report",            
                            ],
                        )
            
                with gr.Row():
                    with gr.Column():
                        diagnosis = gr.CheckboxGroup(
                            label="Diagnosis", 
                            choices=[
                                "Diagnosis",
                            ],
                            value=[
                                "Diagnosis",
                            ],
                        )
                    with gr.Column():
                        treatment = gr.CheckboxGroup(
                            label="Treatment", 
                            choices=[
                                "Interventional Therapy",
                                "Pharmacological Therapy",
                            ],
                            value=[
                                "Interventional Therapy",
                                "Pharmacological Therapy",
                            ],
                        )
                    with gr.Column():
                        patient_outcome = gr.CheckboxGroup(
                            label="Patient Outcome", 
                            choices=[
                                "Patient Outcome",
                            ],
                            value=[
                                "Patient Outcome",
                            ],
                        )               
        
            with gr.Row():
                summarize_btn = gr.Button("Extract")
        
            with gr.Row():
                output_text = gr.HTML()
        
            summarize_btn.click(
                fn=summarize, 
                inputs=[input_text, personal_info, medical_history, clinical_presentation, medical_assessment, diagnosis, treatment, patient_outcome], 
                outputs=output_text, 
                show_progress=True,
            )

        with gr.Tab("Help"):
            gr.Markdown("""## Help
            
### Personal Information        

**Age**: Age of the patient.<br>
**Gender**: Gender of the patient.<br>
**Lifestyle**: Daily habits and activities of the patient (e.g. alcohol consumption, diet, smoking status).<br>
**Social Background**: Social factors of the patient (e.g. housing situation, marital status).<br>

### Medical History

**Personal**: Past medical conditions, previous surgeries or treatments of the patient.<br>
**Family Members**: Relevant medical conditions or genetic disorders in the patient’s family (e.g. cancer, heart disease).<br>

### Clinical Presentation

**Symptoms**: Current symptoms of the patient.<br>
**Comorbid Conditions**: Other medical conditions of the patient that may influence the treatment.<br>

### Medical Assessment

**Diagnostic Procedures**: Description of the diagnostic tests or procedures performed (e.g. X-rays, MRIs)<br>
**Laboratory Results**: Results foom laboratory test (e.g. blood counts, electrolyte levels)<br>
**Pathology Report**: Findings from pathological examinations (e.g. biopsy results)<br>

### Diagnosis

**Diagnosis**: All levels of diagnosis mentioned in the report.<br>

### Treatment

**Interventional Therapy**: Medications prescribed to the patient.<br>
**Pharmacological Therapy**: Information on surgical or non-surgical interventions performed.<br>

### Patient Outcome

**Patient Outcome**: Evaluation of the patient’s health status at the end of treatment.<br>
""")
            
        with gr.Tab("About"):
            gr.Markdown("""## About

LlamaMD is a project developed as part of the "NLP for Social Good" course at TU Berlin.
            
The goal of this project is to perform structured information extraction from clinical reports, helping doctors to have more time for their patients.
                        
The system is based on `meta-llama/Llama-3.2-1B-Instruct`, which has been fine-tuned on the ELMTEX dataset.
""")

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