CraAssitant / app.py
mgbam's picture
Upload 2 files
cd0b15a verified
raw
history blame
25.4 kB
import os
import json
import csv
import asyncio
import xml.etree.ElementTree as ET
from typing import Any, Dict, Optional, Tuple, Union, List
import httpx
import gradio as gr
import torch
from dotenv import load_dotenv
from loguru import logger
from huggingface_hub import login
from openai import OpenAI
from reportlab.pdfgen import canvas
from transformers import (
AutoTokenizer,
AutoModelForSequenceClassification,
MarianMTModel,
MarianTokenizer,
)
import pandas as pd
import altair as alt
import spacy
import spacy.cli
import PyPDF2 # For PDF reading
# Ensure spaCy model is downloaded
try:
nlp = spacy.load("en_core_web_sm")
except OSError:
logger.info("Downloading SpaCy 'en_core_web_sm' model...")
spacy.cli.download("en_core_web_sm")
nlp = spacy.load("en_core_web_sm")
# Logging
logger.add("error_logs.log", rotation="1 MB", level="ERROR")
# Load environment variables
load_dotenv()
HUGGINGFACE_TOKEN = os.getenv("HF_TOKEN")
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
ENTREZ_EMAIL = os.getenv("ENTREZ_EMAIL")
# Basic checks
if not HUGGINGFACE_TOKEN or not OPENAI_API_KEY:
logger.error("Missing Hugging Face or OpenAI credentials.")
raise ValueError("Missing credentials for Hugging Face or OpenAI.")
# API endpoints
PUBMED_SEARCH_URL = "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi"
PUBMED_FETCH_URL = "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi"
EUROPE_PMC_BASE_URL = "https://www.ebi.ac.uk/europepmc/webservices/rest/search"
# Hugging Face login
login(HUGGINGFACE_TOKEN)
# Initialize OpenAI
client = OpenAI(api_key=OPENAI_API_KEY)
# Device setting
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Using device: {device}")
# Model settings
MODEL_NAME = "mgbam/bert-base-finetuned-mgbam"
try:
model = AutoModelForSequenceClassification.from_pretrained(
MODEL_NAME, use_auth_token=HUGGINGFACE_TOKEN
).to(device)
tokenizer = AutoTokenizer.from_pretrained(
MODEL_NAME, use_auth_token=HUGGINGFACE_TOKEN
)
except Exception as e:
logger.error(f"Model load error: {e}")
raise
# Translation model settings
try:
translation_model_name = "Helsinki-NLP/opus-mt-en-fr"
translation_model = MarianMTModel.from_pretrained(
translation_model_name, use_auth_token=HUGGINGFACE_TOKEN
).to(device)
translation_tokenizer = MarianTokenizer.from_pretrained(
translation_model_name, use_auth_token=HUGGINGFACE_TOKEN
)
except Exception as e:
logger.error(f"Translation model load error: {e}")
raise
LANGUAGE_MAP: Dict[str, Tuple[str, str]] = {
"English to French": ("en", "fr"),
"French to English": ("fr", "en"),
}
### Utility Functions ###
def safe_json_parse(text: str) -> Union[Dict, None]:
"""Safely parse JSON string into a Python dictionary."""
try:
return json.loads(text)
except json.JSONDecodeError as e:
logger.error(f"JSON parsing error: {e}")
return None
def parse_pubmed_xml(xml_data: str) -> List[Dict[str, Any]]:
"""Parses PubMed XML data and returns a list of structured articles."""
root = ET.fromstring(xml_data)
articles = []
for article in root.findall(".//PubmedArticle"):
pmid = article.findtext(".//PMID")
title = article.findtext(".//ArticleTitle")
abstract = article.findtext(".//AbstractText")
journal = article.findtext(".//Journal/Title")
pub_date_elem = article.find(".//JournalIssue/PubDate")
pub_date = None
if pub_date_elem is not None:
year = pub_date_elem.findtext("Year")
month = pub_date_elem.findtext("Month")
day = pub_date_elem.findtext("Day")
if year and month and day:
pub_date = f"{year}-{month}-{day}"
else:
pub_date = year
articles.append({
"PMID": pmid,
"Title": title,
"Abstract": abstract,
"Journal": journal,
"PublicationDate": pub_date,
})
return articles
### Async Functions for Europe PMC ###
async def fetch_articles_by_nct_id(nct_id: str) -> Dict[str, Any]:
params = {"query": nct_id, "format": "json"}
async with httpx.AsyncClient() as client_http:
try:
response = await client_http.get(EUROPE_PMC_BASE_URL, params=params)
response.raise_for_status()
return response.json()
except Exception as e:
logger.error(f"Error fetching articles for {nct_id}: {e}")
return {"error": str(e)}
async def fetch_articles_by_query(query_params: str) -> Dict[str, Any]:
parsed_params = safe_json_parse(query_params)
if not parsed_params or not isinstance(parsed_params, dict):
return {"error": "Invalid JSON."}
query_string = " AND ".join(f"{k}:{v}" for k, v in parsed_params.items())
params = {"query": query_string, "format": "json"}
async with httpx.AsyncClient() as client_http:
try:
response = await client_http.get(EUROPE_PMC_BASE_URL, params=params)
response.raise_for_status()
return response.json()
except Exception as e:
logger.error(f"Error fetching articles: {e}")
return {"error": str(e)}
### PubMed Integration ###
async def fetch_pubmed_by_query(query_params: str) -> Dict[str, Any]:
parsed_params = safe_json_parse(query_params)
if not parsed_params or not isinstance(parsed_params, dict):
return {"error": "Invalid JSON for PubMed."}
search_params = {
"db": "pubmed",
"retmode": "json",
"email": ENTREZ_EMAIL,
"retmax": parsed_params.get("retmax", "10"),
"term": parsed_params.get("term", ""),
}
async with httpx.AsyncClient() as client_http:
try:
search_response = await client_http.get(PUBMED_SEARCH_URL, params=search_params)
search_response.raise_for_status()
search_data = search_response.json()
id_list = search_data.get("esearchresult", {}).get("idlist", [])
if not id_list:
return {"result": ""}
fetch_params = {
"db": "pubmed",
"id": ",".join(id_list),
"retmode": "xml",
"email": ENTREZ_EMAIL,
}
fetch_response = await client_http.get(PUBMED_FETCH_URL, params=fetch_params)
fetch_response.raise_for_status()
return {"result": fetch_response.text}
except Exception as e:
logger.error(f"Error fetching PubMed articles: {e}")
return {"error": str(e)}
### Crossref Integration ###
async def fetch_crossref_by_query(query_params: str) -> Dict[str, Any]:
parsed_params = safe_json_parse(query_params)
if not parsed_params or not isinstance(parsed_params, dict):
return {"error": "Invalid JSON for Crossref."}
CROSSREF_API_URL = "https://api.crossref.org/works"
async with httpx.AsyncClient() as client_http:
try:
response = await client_http.get(CROSSREF_API_URL, params=parsed_params)
response.raise_for_status()
return response.json()
except Exception as e:
logger.error(f"Error fetching Crossref data: {e}")
return {"error": str(e)}
### Core Functions ###
def summarize_text(text: str) -> str:
"""Summarize text using OpenAI."""
if not text.strip():
return "No text provided for summarization."
try:
response = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[{"role": "user", "content": f"Summarize the following clinical data:\n{text}"}],
max_tokens=200,
temperature=0.7,
)
return response.choices[0].message.content.strip()
except Exception as e:
logger.error(f"Summarization Error: {e}")
return "Summarization failed."
def predict_outcome(text: str) -> Union[Dict[str, float], str]:
"""Predict outcomes (classification) using a fine-tuned model."""
if not text.strip():
return "No text provided for prediction."
try:
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
inputs = {k: v.to(device) for k, v in inputs.items()}
with torch.no_grad():
outputs = model(**inputs)
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)[0]
return {f"Label {i+1}": float(prob.item()) for i, prob in enumerate(probabilities)}
except Exception as e:
logger.error(f"Prediction Error: {e}")
return "Prediction failed."
def generate_report(text: str, filename: str = "clinical_report.pdf") -> Optional[str]:
"""Generate a PDF report from the given text."""
try:
if not text.strip():
logger.warning("No text provided for the report.")
c = canvas.Canvas(filename)
c.drawString(100, 750, "Clinical Research Report")
lines = text.split("\n")
y = 730
for line in lines:
if y < 50:
c.showPage()
y = 750
c.drawString(100, y, line)
y -= 15
c.save()
logger.info(f"Report generated: {filename}")
return filename
except Exception as e:
logger.error(f"Report Generation Error: {e}")
return None
def visualize_predictions(predictions: Dict[str, float]) -> Optional[alt.Chart]:
"""Visualize model prediction probabilities using Altair."""
try:
data = pd.DataFrame(list(predictions.items()), columns=["Label", "Probability"])
chart = (
alt.Chart(data)
.mark_bar()
.encode(
x=alt.X("Label:N", sort=None),
y="Probability:Q",
tooltip=["Label", "Probability"],
)
.properties(title="Prediction Probabilities", width=500, height=300)
)
return chart
except Exception as e:
logger.error(f"Visualization Error: {e}")
return None
def translate_text(text: str, translation_option: str) -> str:
"""Translate text between English and French."""
if not text.strip():
return "No text provided for translation."
try:
if translation_option not in LANGUAGE_MAP:
return "Unsupported translation option."
inputs = translation_tokenizer(text, return_tensors="pt", padding=True).to(device)
translated_tokens = translation_model.generate(**inputs)
return translation_tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
except Exception as e:
logger.error(f"Translation Error: {e}")
return "Translation failed."
def perform_named_entity_recognition(text: str) -> str:
"""Perform Named Entity Recognition (NER) using spaCy."""
if not text.strip():
return "No text provided for NER."
try:
doc = nlp(text)
entities = [(ent.text, ent.label_) for ent in doc.ents]
if not entities:
return "No named entities found."
return "\n".join(f"{ent_text} -> {ent_label}" for ent_text, ent_label in entities)
except Exception as e:
logger.error(f"NER Error: {e}")
return "Named Entity Recognition failed."
### Enhanced EDA ###
def perform_enhanced_eda(df: pd.DataFrame) -> Tuple[str, Optional[alt.Chart], Optional[alt.Chart]]:
"""
Perform a more advanced EDA given a DataFrame:
- Show dataset info (columns, shape, numeric summary).
- Generate a correlation heatmap (for numeric columns).
- Generate distribution plots (histograms) for numeric columns.
Returns (text_summary, correlation_chart, distribution_chart).
"""
try:
# Basic info
columns_info = f"Columns: {list(df.columns)}"
shape_info = f"Shape: {df.shape[0]} rows x {df.shape[1]} columns"
# Use describe with "include='all'" to show all columns summary
with pd.option_context("display.max_colwidth", 200, "display.max_rows", None):
describe_info = df.describe(include="all").to_string()
summary_text = (
f"--- Enhanced EDA Summary ---\n"
f"{columns_info}\n{shape_info}\n\n"
f"Summary Statistics:\n{describe_info}\n"
)
# Correlation heatmap
numeric_cols = df.select_dtypes(include="number")
corr_chart = None
if numeric_cols.shape[1] >= 2:
corr = numeric_cols.corr()
corr_melted = corr.reset_index().melt(id_vars="index")
corr_melted.columns = ["Feature1", "Feature2", "Correlation"]
corr_chart = (
alt.Chart(corr_melted)
.mark_rect()
.encode(
x="Feature1:O",
y="Feature2:O",
color="Correlation:Q",
tooltip=["Feature1", "Feature2", "Correlation"]
)
.properties(width=400, height=400, title="Correlation Heatmap")
)
# Distribution plots (histograms) for numeric columns
distribution_chart = None
if numeric_cols.shape[1] >= 1:
df_long = numeric_cols.melt(var_name='Column', value_name='Value')
distribution_chart = (
alt.Chart(df_long)
.mark_bar()
.encode(
alt.X("Value:Q", bin=alt.Bin(maxbins=30)),
alt.Y('count()'),
alt.Facet('Column:N', columns=2),
tooltip=["Value"]
)
.properties(
title='Distribution of Numeric Columns',
width=300,
height=200
)
.interactive()
)
return summary_text, corr_chart, distribution_chart
except Exception as e:
logger.error(f"Enhanced EDA Error: {e}")
return f"Enhanced EDA failed: {e}", None, None
### File Handling ###
def read_uploaded_file(uploaded_file: Optional[gr.File]) -> str:
"""
Reads the content of an uploaded file (txt, csv, xls, xlsx, pdf).
Returns the extracted text or CSV-like content.
"""
if uploaded_file is None:
return ""
file_name = uploaded_file.name
file_ext = os.path.splitext(file_name)[1].lower()
try:
# For text
if file_ext == ".txt":
return uploaded_file.read().decode("utf-8")
# For CSV
elif file_ext == ".csv":
return uploaded_file.read().decode("utf-8")
# For Excel
elif file_ext in [".xls", ".xlsx"]:
# We'll just return empty here and parse it later into a DataFrame
# because we can read the binary directly into pd.read_excel().
# Or store as bytes for later use in EDA.
return "EXCEL_FILE_PLACEHOLDER" # We'll handle it differently in EDA step
# For PDF
elif file_ext == ".pdf":
pdf_reader = PyPDF2.PdfReader(uploaded_file)
text_content = []
for page in pdf_reader.pages:
text_content.append(page.extract_text())
return "\n".join(text_content)
else:
return f"Unsupported file format: {file_ext}"
except Exception as e:
logger.error(f"File read error: {e}")
return f"Error reading file: {e}"
def parse_excel_file(uploaded_file) -> pd.DataFrame:
"""
Parse an Excel file into a pandas DataFrame.
We assume the user wants the first sheet or we can guess.
"""
try:
# For Excel, we can do:
df = pd.read_excel(uploaded_file, engine="openpyxl")
return df
except Exception as e:
logger.error(f"Excel parsing error: {e}")
raise
def parse_csv_content(csv_content: str) -> pd.DataFrame:
"""
Attempt to parse CSV content with both utf-8 and utf-8-sig to handle BOM issues.
"""
from io import StringIO
errors = []
for encoding_try in ["utf-8", "utf-8-sig"]:
try:
df = pd.read_csv(StringIO(csv_content), encoding=encoding_try)
return df
except Exception as e:
errors.append(f"Encoding {encoding_try} failed: {e}")
error_msg = "Could not parse CSV content.\n" + "\n".join(errors)
logger.error(error_msg)
raise ValueError(error_msg)
### Gradio Interface ###
with gr.Blocks() as demo:
gr.Markdown("# ✨ Advanced Clinical Research Assistant with Enhanced EDA ✨")
gr.Markdown("""
Welcome to the **Enhanced** AI-Powered Clinical Assistant!
- **Summarize** large blocks of clinical text.
- **Predict** outcomes with a fine-tuned model.
- **Translate** text between English & French.
- **Perform Named Entity Recognition** with spaCy.
- **Fetch** from PubMed, Crossref, Europe PMC.
- **Generate** professional PDF reports.
- **Perform Enhanced EDA** on CSV/Excel data with correlation heatmaps & distribution plots.
""")
# Inputs
with gr.Row():
text_input = gr.Textbox(label="Input Text", lines=5, placeholder="Enter clinical text or query...")
file_input = gr.File(
label="Upload File (txt/csv/xls/xlsx/pdf)",
file_types=[".txt", ".csv", ".xls", ".xlsx", ".pdf"]
)
action = gr.Radio(
[
"Summarize",
"Predict Outcome",
"Generate Report",
"Translate",
"Perform Named Entity Recognition",
"Perform Enhanced EDA",
"Fetch Clinical Studies",
"Fetch PubMed Articles (Legacy)",
"Fetch PubMed by Query",
"Fetch Crossref by Query",
],
label="Select an Action",
)
translation_option = gr.Dropdown(
choices=list(LANGUAGE_MAP.keys()),
label="Translation Option",
value="English to French"
)
query_params_input = gr.Textbox(
label="Query Parameters (JSON Format)",
placeholder='{"term": "cancer", "retmax": "5"}'
)
nct_id_input = gr.Textbox(label="NCT ID for Article Search")
report_filename_input = gr.Textbox(
label="Report Filename",
placeholder="clinical_report.pdf",
value="clinical_report.pdf"
)
export_format = gr.Dropdown(["None", "CSV", "JSON"], label="Export Format")
# Outputs
output_text = gr.Textbox(label="Output", lines=10)
with gr.Row():
output_chart = gr.Plot(label="Visualization 1")
output_chart2 = gr.Plot(label="Visualization 2")
output_file = gr.File(label="Generated File")
submit_button = gr.Button("Submit")
# Async function for handling actions
async def handle_action(
action: str,
text: str,
file_up: gr.File,
translation_opt: str,
query_params: str,
nct_id: str,
report_filename: str,
export_format: str
) -> Tuple[Optional[str], Optional[Any], Optional[Any], Optional[str]]:
# Read the uploaded file
file_content = read_uploaded_file(file_up)
combined_text = (text + "\n" + file_content).strip() if file_content else text
# Branch by action
if action == "Summarize":
return summarize_text(combined_text), None, None, None
elif action == "Predict Outcome":
predictions = predict_outcome(combined_text)
if isinstance(predictions, dict):
chart = visualize_predictions(predictions)
return json.dumps(predictions, indent=2), chart, None, None
return predictions, None, None, None
elif action == "Generate Report":
file_path = generate_report(combined_text, filename=report_filename)
msg = f"Report generated: {file_path}" if file_path else "Report generation failed."
return msg, None, None, file_path
elif action == "Translate":
return translate_text(combined_text, translation_opt), None, None, None
elif action == "Perform Named Entity Recognition":
ner_result = perform_named_entity_recognition(combined_text)
return ner_result, None, None, None
elif action == "Perform Enhanced EDA":
# We expect the user to either upload a CSV or Excel, or paste CSV content.
if file_up is None and not combined_text:
return "No data provided for EDA.", None, None, None
# If Excel was uploaded
if file_up and file_up.name.lower().endswith((".xls", ".xlsx")):
try:
df_excel = parse_excel_file(file_up)
eda_summary, corr_chart, dist_chart = perform_enhanced_eda(df_excel)
return eda_summary, corr_chart, dist_chart, None
except Exception as e:
return f"Excel EDA failed: {e}", None, None, None
# If CSV was uploaded
if file_up and file_up.name.lower().endswith(".csv"):
try:
df_csv = parse_csv_content(file_content)
eda_summary, corr_chart, dist_chart = perform_enhanced_eda(df_csv)
return eda_summary, corr_chart, dist_chart, None
except Exception as e:
return f"CSV EDA failed: {e}", None, None, None
# If user just pasted CSV content (no file)
if not file_up and "," in combined_text:
try:
df_csv = parse_csv_content(combined_text)
eda_summary, corr_chart, dist_chart = perform_enhanced_eda(df_csv)
return eda_summary, corr_chart, dist_chart, None
except Exception as e:
return f"CSV EDA failed: {e}", None, None, None
# Otherwise, not supported
return "No valid CSV/Excel data found for EDA.", None, None, None
elif action == "Fetch Clinical Studies":
if nct_id:
result = await fetch_articles_by_nct_id(nct_id)
elif query_params:
result = await fetch_articles_by_query(query_params)
else:
return "Provide either an NCT ID or valid query parameters.", None, None, None
articles = result.get("resultList", {}).get("result", [])
if not articles:
return "No articles found.", None, None, None
formatted_results = "\n\n".join(
f"Title: {a.get('title')}\nJournal: {a.get('journalTitle')} ({a.get('pubYear')})"
for a in articles
)
return formatted_results, None, None, None
elif action in ["Fetch PubMed Articles (Legacy)", "Fetch PubMed by Query"]:
pubmed_result = await fetch_pubmed_by_query(query_params)
xml_data = pubmed_result.get("result")
if xml_data:
articles = parse_pubmed_xml(xml_data)
if not articles:
return "No articles found.", None, None, None
formatted = "\n\n".join(
f"{a['Title']} - {a['Journal']} ({a['PublicationDate']})"
for a in articles if a['Title']
)
return formatted if formatted else "No articles found.", None, None, None
return "No articles found or error fetching data.", None, None, None
elif action == "Fetch Crossref by Query":
crossref_result = await fetch_crossref_by_query(query_params)
items = crossref_result.get("message", {}).get("items", [])
if not items:
return "No results found.", None, None, None
formatted = "\n\n".join(
f"Title: {item.get('title', ['No title'])[0]}, DOI: {item.get('DOI')}"
for item in items
)
return formatted, None, None, None
return "Invalid action.", None, None, None
submit_button.click(
handle_action,
inputs=[
action,
text_input,
file_input,
translation_option,
query_params_input,
nct_id_input,
report_filename_input,
export_format,
],
outputs=[output_text, output_chart, output_chart2, output_file],
)
# Launch the Gradio app
demo.launch(server_name="0.0.0.0", server_port=7860, share=True)