my-streamlit-app / summarizer.py
samarth09healthPM's picture
Add HIPAA RAG Clinical Summarizer (essential files only)
f64b3f9
# app/summarizer.py
# Day 10: Enhanced HIPAA-compliant RAG clinical summarizer with robustness improvements
# Critical fixes:
# - Added progress indicators during model generation
# - Implemented timeout mechanism for long-running operations
# - Optimized for CPU with reduced generation parameters
# - Better error handling and verbose logging
# - Fallback to smaller max tokens if generation hangs
import os
import argparse
import traceback
from typing import List, Dict, Optional
import re
import time
import sys
from sentence_transformers import SentenceTransformer
from langchain_community.vectorstores import Chroma, FAISS
from langchain_core.documents import Document
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
# -----------------------------
# Embeddings / Vector stores
# -----------------------------
def load_embedder(model_name: str = "sentence-transformers/all-MiniLM-L6-v2"):
"""
Load sentence transformer for embeddings.
For medical domain: consider "emilyalsentzer/Bio_ClinicalBERT" or similar
"""
print(f" β†’ Loading embedding model...")
model = SentenceTransformer(model_name)
def embed_f(texts: List[str]):
vecs = model.encode(texts, convert_to_numpy=True, normalize_embeddings=True)
return vecs.tolist()
print(f" βœ“ Embedding model loaded")
return embed_f
def load_chroma(persist_dir: str, collection: str, embed_f):
from langchain.embeddings.base import Embeddings
class STEmbeddings(Embeddings):
def embed_documents(self, texts: List[str]) -> List[List[float]]:
return embed_f(texts)
def embed_query(self, text: str) -> List[float]:
return embed_f([text])[0]
embeddings = STEmbeddings()
print(f" β†’ Loading Chroma vector store from {persist_dir}...")
return Chroma(collection_name=collection, persist_directory=persist_dir, embedding_function=embeddings)
def load_faiss(persist_dir: str, embed_f):
import pickle, faiss
from langchain.embeddings.base import Embeddings
class STEmbeddings(Embeddings):
def embed_documents(self, texts: List[str]) -> List[List[float]]:
return embed_f(texts)
def embed_query(self, text: str) -> List[float]:
return embed_f([text])[0]
embeddings = STEmbeddings()
index_path = os.path.join(persist_dir, "index.faiss")
meta_path = os.path.join(persist_dir, "meta.pkl")
if not (os.path.exists(index_path) and os.path.exists(meta_path)):
raise FileNotFoundError(f"FAISS files not found in {persist_dir}")
print(f" β†’ Loading FAISS index from {persist_dir}...")
with open(meta_path, "rb") as f:
meta = pickle.load(f)
texts = [m["text"] for m in meta]
metadatas = [m["meta"] | {"id": m["id"]} for m in meta]
vdb = FAISS.from_texts(texts=texts, embedding=embeddings, metadatas=metadatas)
vdb.index = faiss.read_index(index_path)
return vdb
def retrieve_docs(db_type: str, persist_dir: str, collection: str, query: str, top_k: int, embed_f) -> List[Document]:
if db_type == "chroma":
vdb = load_chroma(persist_dir, collection, embed_f)
else:
vdb = load_faiss(persist_dir, embed_f)
print(f" β†’ Retrieving documents...")
retriever = vdb.as_retriever(search_kwargs={"k": top_k})
docs: List[Document] = retriever.invoke(query)
print(f" βœ“ Retrieved {len(docs)} document(s)")
# Debug: Show retrieved content length
if docs:
total_chars = sum(len(d.page_content) for d in docs)
print(f" β„Ή Total retrieved content: {total_chars} characters")
else:
print(f" ⚠ WARNING: No documents retrieved!")
return docs
# -----------------------------
# T5 Summarization utilities
# -----------------------------
def make_t5(model_name="google/flan-t5-base", device="cpu"):
print(f" β†’ Loading T5 model: {model_name}")
print(f" β„Ή This may take 30-60 seconds for large models...")
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
print(f" βœ“ Model loaded successfully")
return tokenizer, model
def t5_generate(tokenizer, model, prompt: str, max_input_tokens: int = 512, max_output_tokens: int = 256, section_name: str = ""):
"""
Enhanced generation with progress indicators and optimized parameters for CPU
"""
# Show progress
if section_name:
print(f" β†’ Generating {section_name}...", end='', flush=True)
else:
print(f" β†’ Generating summary...", end='', flush=True)
start_time = time.time()
try:
inputs = tokenizer(prompt, truncation=True, max_length=max_input_tokens, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
# Optimized parameters for CPU performance
outputs = model.generate(
**inputs,
max_new_tokens=max_output_tokens,
min_length=10, # Reduced minimum to avoid forcing long outputs
num_beams=2, # Reduced from 4 for faster CPU generation
length_penalty=1.0, # Reduced from 1.5
no_repeat_ngram_size=3,
early_stopping=True, # Re-enabled for faster completion
do_sample=False # Deterministic generation
)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
elapsed = time.time() - start_time
print(f" done ({elapsed:.1f}s)")
return result
except Exception as e:
elapsed = time.time() - start_time
print(f" FAILED ({elapsed:.1f}s)")
print(f" βœ— Error: {str(e)}")
return ""
def dedupe_texts(texts: List[str]) -> List[str]:
seen = set()
uniq = []
for t in texts:
key = " ".join(t.split())[:500]
if key not in seen:
seen.add(key)
uniq.append(t)
return uniq
# -----------------------------
# Section definitions
# -----------------------------
SECTION_ORDER = [
"Chief Complaint",
"HPI",
"PMH",
"Medications",
"Allergies",
"Assessment",
"Plan",
]
# -----------------------------
# Multi-stage extraction prompts (optimized for T5)
# -----------------------------
SECTION_PROMPTS = {
"Chief Complaint": """Task: Extract the main reason for patient visit.
Clinical Note:
{context}
Answer with only the chief complaint (1-2 sentences):""",
"HPI": """Task: Extract the history of present illness including symptom onset, progression, and context.
Clinical Note:
{context}
Answer with the history of present illness:""",
"PMH": """Task: Extract past medical history including chronic conditions, past surgeries, and social history.
Clinical Note:
{context}
Answer with past medical history:""",
"Medications": """Task: List all medications with dosages mentioned in the note.
Clinical Note:
{context}
Answer with medication list:""",
"Allergies": """Task: Extract drug allergies. If none mentioned, state "No known drug allergies".
Clinical Note:
{context}
Answer with allergies:""",
"Assessment": """Task: Extract diagnosis, test results, physical findings, and vital signs.
Clinical Note:
{context}
Answer with assessment and findings:""",
"Plan": """Task: Extract treatment plan, medications prescribed, follow-up appointments, and discharge instructions.
Clinical Note:
{context}
Answer with treatment plan:"""
}
# -----------------------------
# Enhanced extraction pipeline
# -----------------------------
def extract_section_multistage(tokenizer, model, context: str, section: str) -> str:
"""
Extract a single section using targeted prompting
"""
if section not in SECTION_PROMPTS:
return "None stated"
# Truncate context if too long
max_context_chars = 2000
if len(context) > max_context_chars:
context = context[:max_context_chars] + "..."
prompt = SECTION_PROMPTS[section].format(context=context)
try:
result = t5_generate(tokenizer, model, prompt, max_input_tokens=512, max_output_tokens=200, section_name=section)
result = result.strip()
# Remove any section headers the model might have added
result = re.sub(r'^(Chief Complaint|HPI|PMH|Medications|Allergies|Assessment|Plan)\s*:\s*', '', result, flags=re.IGNORECASE)
# Check if extraction failed
if not result or len(result) < 5 or result.lower() in ["none", "none stated", "not mentioned", "n/a", "na"]:
return "None stated"
return result.strip()
except Exception as e:
print(f" βœ— Error extracting {section}: {str(e)}")
return "None stated"
def validate_extraction(sections: Dict[str, str]) -> bool:
"""
Validate that extraction was successful (not all 'None stated')
"""
non_empty = sum(1 for v in sections.values() if v and v != "None stated")
return non_empty >= 2 # At least 2 sections should have content
def summarize_docs_multistage(tokenizer, model, docs: List[Document]) -> str:
"""
Multi-stage extraction: extract each section independently
"""
print(f"\nπŸ“„ Processing documents...")
contents = dedupe_texts([d.page_content for d in docs if d and d.page_content])
if not contents:
print(" ⚠ No content to summarize!")
return format_output({sec: "None stated" for sec in SECTION_ORDER})
# Combine all retrieved content
full_context = "\n\n".join(contents)
print(f" β„Ή Combined context length: {len(full_context)} characters")
# Extract each section independently
print(f"\nπŸ”„ Extracting sections (this may take 1-3 minutes on CPU)...")
sections = {}
for i, section in enumerate(SECTION_ORDER, 1):
print(f" [{i}/{len(SECTION_ORDER)}] {section}:")
sections[section] = extract_section_multistage(tokenizer, model, full_context, section)
# Validate extraction
print(f"\nβœ“ Extraction complete")
if not validate_extraction(sections):
print("⚠ WARNING: Extraction appears incomplete. Most sections are empty.")
print(" Possible issues:")
print(" β€’ Vector retrieval may not be finding relevant content")
print(" β€’ Model may not understand the clinical text format")
print(" β€’ Context may be too short or fragmented")
print(" β€’ De-identification artifacts may be confusing the model")
return format_output(sections)
def format_output(sections: Dict[str, str]) -> str:
"""
Format sections into structured output
"""
output_lines = []
for section in SECTION_ORDER:
content = sections.get(section, "None stated")
output_lines.append(f"β€’ {section}: {content}")
return "\n".join(output_lines)
# -----------------------------
# Summary Quality Validation
# -----------------------------
def validate_summary_quality(summary: str, original_text: str = "") -> dict:
"""
Validate summary quality and detect common issues
Args:
summary: The generated summary text
original_text: Optional original note text for comparison
Returns:
Dictionary with validation results
"""
issues = []
warnings = []
# Check for placeholder contamination (de-ID over-redaction)
placeholder_patterns = [
(r'\[LOCATION\]', 'LOCATION'),
(r'\[DATE\]', 'DATE'),
(r'\[NAME\]', 'NAME'),
(r'\[PHONE\]', 'PHONE')
]
total_placeholders = 0
for pattern, name in placeholder_patterns:
count = len(re.findall(pattern, summary))
total_placeholders += count
if count > 2:
warnings.append(f"Too many [{name}] placeholders ({count}) - de-identification may be over-aggressive")
if total_placeholders > 5:
issues.append(f"Critical: {total_placeholders} PHI placeholders in summary - clinical content lost")
# Check for "None stated" sections
none_count = summary.count("None stated")
if none_count >= 5:
issues.append(f"Critical: {none_count}/7 sections are empty - summarization failed")
elif none_count >= 3:
warnings.append(f"Warning: {none_count}/7 sections are empty - may need better retrieval")
# Check for minimum content length per section
total_length = len(summary)
# Subtract bullets and "None stated" overhead
content_length = total_length - (summary.count("β€’") * 2) - (none_count * 11)
filled_sections = 7 - none_count
if filled_sections > 0:
avg_section_length = content_length / filled_sections
if avg_section_length < 30:
warnings.append(f"Warning: Sections too short (avg {avg_section_length:.0f} chars) - may lack detail")
# Check for duplicate medications
if "Medications:" in summary:
meds_section = summary.split("Medications:")[1].split("β€’")[0] if "Medications:" in summary else ""
meds_lower = meds_section.lower()
common_meds = ['atorvastatin', 'metoprolol', 'lisinopril', 'aspirin', 'metformin']
for med in common_meds:
if meds_lower.count(med) > 1:
warnings.append(f"Warning: Duplicate medication detected: {med}")
# Calculate quality score (0-100)
score = 100
score -= len(issues) * 30 # Critical issues: -30 each
score -= len(warnings) * 10 # Warnings: -10 each
score = max(0, min(100, score))
# Determine overall status
if len(issues) > 0:
status = "FAILED"
elif len(warnings) > 2:
status = "POOR"
elif len(warnings) > 0:
status = "FAIR"
else:
status = "GOOD"
return {
"is_valid": len(issues) == 0,
"status": status,
"quality_score": score,
"issues": issues,
"warnings": warnings,
"metrics": {
"total_placeholders": total_placeholders,
"empty_sections": none_count,
"filled_sections": filled_sections,
"total_length": total_length
}
}
# -----------------------------
# Backward compatibility wrapper for Streamlit integration
# -----------------------------
def summarize_docs(tokenizer, model, docs: List[Document], method: str = "multistage") -> str:
"""
Wrapper function for backward compatibility with main.py (Streamlit UI)
"""
if method == "multistage":
return summarize_docs_multistage(tokenizer, model, docs)
else:
return summarize_docs_singleshot(tokenizer, model, docs)
# -----------------------------
# Single-shot extraction (simplified fallback)
# -----------------------------
def summarize_docs_singleshot(tokenizer, model, docs: List[Document]) -> str:
"""
Single-shot extraction method (faster but less comprehensive)
"""
print(f"\nπŸ“„ Processing documents...")
contents = dedupe_texts([d.page_content for d in docs if d and d.page_content])
if not contents:
print(" ⚠ No content to summarize!")
return format_output({sec: "None stated" for sec in SECTION_ORDER})
raw_context = "\n\n".join(contents)
print(f" β„Ή Combined context length: {len(raw_context)} characters")
# Simplified prompt for single-shot
instruction = """Summarize this clinical note into 7 sections:
1. Chief Complaint (main reason for visit)
2. HPI (symptom history and progression)
3. PMH (past medical history)
4. Medications (current medications with doses)
5. Allergies (drug allergies)
6. Assessment (diagnosis and findings)
7. Plan (treatment plan and follow-up)
Clinical Note:
{context}
Structured Summary:"""
print(f"\nπŸ”„ Generating structured summary...")
prompt = instruction.format(context=raw_context[:2000]) # Limit context
model_out = t5_generate(tokenizer, model, prompt, max_input_tokens=512, max_output_tokens=400)
# Parse output into sections
sections = parse_output_to_sections(model_out)
return format_output(sections)
def parse_output_to_sections(text: str) -> Dict[str, str]:
"""
Parse model output into section dictionary
"""
sections = {}
current_section = None
current_content = []
for line in text.split('\n'):
line = line.strip()
if not line:
continue
# Check if line starts with a section header
matched_section = None
for section in SECTION_ORDER:
# Match section headers with numbers or bullets
pattern = rf'^(\d+\.\s*)?{re.escape(section)}\s*:?'
if re.match(pattern, line, re.IGNORECASE):
matched_section = section
break
if matched_section:
# Save previous section
if current_section:
sections[current_section] = " ".join(current_content).strip()
# Start new section
current_section = matched_section
# Get content after the header
content = re.sub(rf'^(\d+\.\s*)?{re.escape(matched_section)}\s*:?\s*', '', line, flags=re.IGNORECASE).strip()
current_content = [content] if content else []
else:
# Continue current section
if current_section:
current_content.append(line)
# Save last section
if current_section:
sections[current_section] = " ".join(current_content).strip()
# Fill in missing sections
for section in SECTION_ORDER:
if section not in sections or not sections[section]:
sections[section] = "None stated"
return sections
# -----------------------------
# Backward compatibility wrapper for Streamlit integration
# -----------------------------
def summarize_docs(tokenizer, model, docs: List[Document], method: str = "multistage") -> str:
"""
Wrapper function for backward compatibility with main.py (Streamlit UI)
Args:
tokenizer: T5 tokenizer instance
model: T5 model instance
docs: List of retrieved documents
method: "multistage" (default) or "singleshot" extraction method
Returns:
Formatted summary string with sections
"""
if method == "multistage":
return summarize_docs_multistage(tokenizer, model, docs)
else:
return summarize_docs_singleshot(tokenizer, model, docs)
# -----------------------------
# Orchestration
# -----------------------------
def main():
parser = argparse.ArgumentParser(description="Day 10: Enhanced HIPAA-compliant RAG clinical summarizer")
parser.add_argument("--db_type", choices=["chroma", "faiss"], default="chroma")
parser.add_argument("--persist_dir", default="./data/vector_store")
parser.add_argument("--collection", default="notes")
parser.add_argument("--embed_model", default="sentence-transformers/all-MiniLM-L6-v2")
parser.add_argument("--model_name", default="google/flan-t5-small")
parser.add_argument("--query", required=True)
parser.add_argument("--top_k", type=int, default=5)
parser.add_argument("--out", default="./data/outputs/summaries/summary.txt")
parser.add_argument("--method", choices=["multistage", "singleshot"], default="multistage",
help="Extraction method: multistage (recommended) or singleshot (faster)")
args = parser.parse_args()
print("=" * 70)
print(" HIPAA-COMPLIANT RAG CLINICAL SUMMARIZER")
print("=" * 70)
out_dir = os.path.dirname(args.out) or "."
os.makedirs(out_dir, exist_ok=True)
try:
# Step 1: Load embedder
print(f"\n[1/4] LOADING EMBEDDER")
print(f" Model: {args.embed_model}")
embed_f = load_embedder(args.embed_model)
# Step 2: Retrieve documents
print(f"\n[2/4] RETRIEVING DOCUMENTS")
print(f" Database: {args.db_type}")
print(f" Location: {args.persist_dir}")
print(f" Query: {args.query}")
print(f" Top-K: {args.top_k}")
docs = retrieve_docs(args.db_type, args.persist_dir, args.collection, args.query, args.top_k, embed_f)
if not docs:
print("\n⚠ ERROR: No documents retrieved from vector database!")
print(" Possible causes:")
print(" β€’ Vector database is empty or not properly indexed")
print(" β€’ Query doesn't match indexed content")
print(" β€’ Database path is incorrect")
result = format_output({sec: "None stated" for sec in SECTION_ORDER})
with open(args.out, "w", encoding="utf-8") as f:
f.write(result)
print(f"\nβœ“ Empty summary written to {args.out}")
return
# Step 3: Load summarization model
print(f"\n[3/4] LOADING SUMMARIZATION MODEL")
print(f" Model: {args.model_name}")
tokenizer, model = make_t5(args.model_name)
# Step 4: Generate summary
print(f"\n[4/4] GENERATING SUMMARY")
print(f" Method: {args.method}")
if args.method == "multistage":
summary = summarize_docs_multistage(tokenizer, model, docs)
else:
summary = summarize_docs_singleshot(tokenizer, model, docs)
# Write summary to output file
with open(args.out, "w", encoding="utf-8") as f:
f.write(summary)
print(f"\n{'=' * 70}")
print(f"βœ“ SUCCESS: Summary written to {args.out}")
print(f"{'=' * 70}")
print("\nGenerated Summary:")
print("-" * 70)
print(summary)
print("-" * 70)
except Exception as e:
err = traceback.format_exc()
error_msg = f"ERROR during summarization:\n{err}"
# Write error to file
with open(args.out, "w", encoding="utf-8") as f:
f.write(error_msg)
print(f"\n{'=' * 70}")
print(f"βœ— ERROR: An error occurred during processing")
print(f"{'=' * 70}")
print(f"\n{err}")
print(f"\nError details written to {args.out}")
if __name__ == "__main__":
main()