TxAgent-Api / app.py
Ali2206's picture
Update app.py
f275c80 verified
raw
history blame
23.8 kB
# app.py (in TxAgent-API)
import os
import sys
import json
import logging
import re
import hashlib
import io
import base64
from datetime import datetime
from typing import List, Dict, Optional, Tuple
from enum import Enum
from fastapi import FastAPI, HTTPException, UploadFile, File, Query, Form, Depends
from fastapi.responses import StreamingResponse, JSONResponse
from fastapi.middleware.cors import CORSMiddleware
from fastapi.security import OAuth2PasswordBearer
from fastapi.encoders import jsonable_encoder
from pydantic import BaseModel
import asyncio
from bson import ObjectId
import speech_recognition as sr
from gtts import gTTS
from pydub import AudioSegment
import PyPDF2
import mimetypes
from docx import Document
from jose import JWTError, jwt
from txagent.txagent import TxAgent
from db.mongo import get_mongo_client
# Logging
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger("TxAgentAPI")
# App
app = FastAPI(title="TxAgent API", version="2.6.0")
# CORS
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"]
)
# JWT settings (must match CPS-API)
SECRET_KEY = os.getenv("SECRET_KEY", "your-secret-key") # Same as CPS-API
ALGORITHM = "HS256"
# OAuth2 scheme (point to CPS-API's login endpoint)
oauth2_scheme = OAuth2PasswordBearer(tokenUrl="https://rocketfarmstudios-cps-api.hf.space/auth/login")
# Pydantic Models
class ChatRequest(BaseModel):
message: str
temperature: float = 0.7
max_new_tokens: int = 512
history: Optional[List[Dict]] = None
format: Optional[str] = "clean"
class VoiceInputRequest(BaseModel):
audio_format: str = "wav"
language: str = "en-US"
class VoiceOutputRequest(BaseModel):
text: str
language: str = "en"
slow: bool = False
return_format: str = "mp3"
# Enums
class RiskLevel(str, Enum):
NONE = "none"
LOW = "low"
MODERATE = "moderate"
HIGH = "high"
SEVERE = "severe"
# Globals
agent = None
patients_collection = None
analysis_collection = None
alerts_collection = None
# JWT validation
async def get_current_user(token: str = Depends(oauth2_scheme)):
credentials_exception = HTTPException(
status_code=401,
detail="Could not validate credentials",
headers={"WWW-Authenticate": "Bearer"},
)
try:
payload = jwt.decode(token, SECRET_KEY, algorithms=[ALGORITHM])
email: str = payload.get("sub")
if email is None:
raise credentials_exception
except JWTError:
raise credentials_exception
user = await users_collection.find_one({"email": email})
if user is None:
raise credentials_exception
return user
# Helper functions (unchanged from your original code)
def clean_text_response(text: str) -> str:
text = re.sub(r'\n\s*\n', '\n\n', text)
text = re.sub(r'[ ]+', ' ', text)
return text.replace("**", "").replace("__", "").strip()
def extract_section(text: str, heading: str) -> str:
try:
pattern = rf"{re.escape(heading)}:\s*\n(.*?)(?=\n[A-Z][^\n]*:|\Z)"
match = re.search(pattern, text, re.DOTALL | re.IGNORECASE)
return match.group(1).strip() if match else ""
except Exception as e:
logger.error(f"Section extraction failed for heading '{heading}': {e}")
return ""
def structure_medical_response(text: str) -> Dict:
def extract_improved(text: str, heading: str) -> str:
patterns = [
rf"{re.escape(heading)}:\s*\n(.*?)(?=\n\s*\n|\Z)",
rf"\*\*{re.escape(heading)}\*\*:\s*\n(.*?)(?=\n\s*\n|\Z)",
rf"{re.escape(heading)}[\s\-]+(.*?)(?=\n\s*\n|\Z)",
rf"\n{re.escape(heading)}\s*\n(.*?)(?=\n\s*\n|\Z)"
]
for pattern in patterns:
match = re.search(pattern, text, re.DOTALL | re.IGNORECASE)
if match:
content = match.group(1).strip()
content = re.sub(r'^\s*[\-\*]\s*', '', content, flags=re.MULTILINE)
return content
return ""
text = text.replace('**', '').replace('__', '')
return {
"summary": extract_improved(text, "Summary of Patient's Medical History") or
extract_improved(text, "Summarize the patient's medical history"),
"risks": extract_improved(text, "Identify Risks or Red Flags") or
extract_improved(text, "Risks or Red Flags"),
"missed_issues": extract_improved(text, "Missed Diagnoses or Treatments") or
extract_improved(text, "What the doctor might have missed"),
"recommendations": extract_improved(text, "Suggest Next Clinical Steps") or
extract_improved(text, "Suggested Clinical Actions")
}
def detect_suicide_risk(text: str) -> Tuple[RiskLevel, float, List[str]]:
suicide_keywords = [
'suicide', 'suicidal', 'kill myself', 'end my life',
'want to die', 'self-harm', 'self harm', 'hopeless',
'no reason to live', 'plan to die'
]
explicit_mentions = [kw for kw in suicide_keywords if kw in text.lower()]
if not explicit_mentions:
return RiskLevel.NONE, 0.0, []
assessment_prompt = (
"Assess the suicide risk level based on this text. "
"Consider frequency, specificity, and severity of statements. "
"Respond with JSON format: {\"risk_level\": \"low/moderate/high/severe\", "
"\"risk_score\": 0-1, \"factors\": [\"list of risk factors\"]}\n\n"
f"Text to assess:\n{text}"
)
try:
response = agent.chat(
message=assessment_prompt,
history=[],
temperature=0.2,
max_new_tokens=256
)
json_match = re.search(r'\{.*\}', response, re.DOTALL)
if json_match:
assessment = json.loads(json_match.group())
return (
RiskLevel(assessment.get("risk_level", "none").lower()),
float(assessment.get("risk_score", 0)),
assessment.get("factors", [])
)
except Exception as e:
logger.error(f"Error in suicide risk assessment: {e}")
risk_score = min(0.1 * len(explicit_mentions), 0.9)
if risk_score > 0.7:
return RiskLevel.HIGH, risk_score, explicit_mentions
elif risk_score > 0.4:
return RiskLevel.MODERATE, risk_score, explicit_mentions
return RiskLevel.LOW, risk_score, explicit_mentions
async def create_alert(patient_id: str, risk_data: dict):
alert_doc = {
"patient_id": patient_id,
"type": "suicide_risk",
"level": risk_data["level"],
"score": risk_data["score"],
"factors": risk_data["factors"],
"timestamp": datetime.utcnow(),
"acknowledged": False
}
await alerts_collection.insert_one(alert_doc)
logger.warning(f"⚠️ Created suicide risk alert for patient {patient_id}")
def serialize_patient(patient: dict) -> dict:
patient_copy = patient.copy()
if "_id" in patient_copy:
patient_copy["_id"] = str(patient_copy["_id"])
return patient_copy
def compute_patient_data_hash(data: dict) -> str:
serialized = json.dumps(data, sort_keys=True)
return hashlib.sha256(serialized.encode()).hexdigest()
def compute_file_content_hash(file_content: bytes) -> str:
return hashlib.sha256(file_content).hexdigest()
def extract_text_from_pdf(pdf_data: bytes) -> str:
try:
pdf_reader = PyPDF2.PdfReader(io.BytesIO(pdf_data))
text = ""
for page in pdf_reader.pages:
text += page.extract_text() or ""
return clean_text_response(text)
except Exception as e:
logger.error(f"Error extracting text from PDF: {e}")
raise HTTPException(status_code=400, detail="Failed to extract text from PDF")
async def analyze_patient_report(patient_id: Optional[str], report_content: str, file_type: str, file_content: bytes):
identifier = patient_id if patient_id else compute_file_content_hash(file_content)
report_data = {"identifier": identifier, "content": report_content, "file_type": file_type}
report_hash = compute_patient_data_hash(report_data)
logger.info(f"🧾 Analyzing report for identifier: {identifier}")
existing_analysis = await analysis_collection.find_one({"identifier": identifier, "report_hash": report_hash})
if existing_analysis:
logger.info(f"✅ No changes in report data for {identifier}, skipping analysis")
return existing_analysis
prompt = (
"You are a clinical decision support AI. Analyze the following patient report:\n"
"1. Summarize the patient's medical history.\n"
"2. Identify risks or red flags (including mental health and suicide risk).\n"
"3. Highlight missed diagnoses or treatments.\n"
"4. Suggest next clinical steps.\n"
f"\nPatient Report ({file_type}):\n{'-'*40}\n{report_content[:10000]}"
)
raw_response = agent.chat(
message=prompt,
history=[],
temperature=0.7,
max_new_tokens=1024
)
structured_response = structure_medical_response(raw_response)
risk_level, risk_score, risk_factors = detect_suicide_risk(raw_response)
suicide_risk = {
"level": risk_level.value,
"score": risk_score,
"factors": risk_factors
}
analysis_doc = {
"identifier": identifier,
"patient_id": patient_id,
"timestamp": datetime.utcnow(),
"summary": structured_response,
"suicide_risk": suicide_risk,
"raw": raw_response,
"report_hash": report_hash,
"file_type": file_type
}
await analysis_collection.update_one(
{"identifier": identifier, "report_hash": report_hash},
{"$set": analysis_doc},
upsert=True
)
if patient_id and risk_level in [RiskLevel.MODERATE, RiskLevel.HIGH, RiskLevel.SEVERE]:
await create_alert(patient_id, suicide_risk)
logger.info(f"✅ Stored analysis for identifier {identifier}")
return analysis_doc
async def analyze_all_patients():
patients = await patients_collection.find({}).to_list(length=None)
for patient in patients:
await analyze_patient(patient)
await asyncio.sleep(0.1)
async def analyze_patient(patient: dict):
try:
serialized = serialize_patient(patient)
patient_id = serialized.get("fhir_id")
patient_hash = compute_patient_data_hash(serialized)
logger.info(f"🧾 Analyzing patient: {patient_id}")
existing_analysis = await analysis_collection.find_one({"patient_id": patient_id})
if existing_analysis and existing_analysis.get("data_hash") == patient_hash:
logger.info(f"✅ No changes in patient data for {patient_id}, skipping analysis")
return
doc = json.dumps(serialized, indent=2)
message = (
"You are a clinical decision support AI.\n\n"
"Given the patient document below:\n"
"1. Summarize the patient's medical history.\n"
"2. Identify risks or red flags (including mental health and suicide risk).\n"
"3. Highlight missed diagnoses or treatments.\n"
"4. Suggest next clinical steps.\n"
f"\nPatient Document:\n{'-'*40}\n{doc[:10000]}"
)
raw = agent.chat(message=message, history=[], temperature=0.7, max_new_tokens=1024)
structured = structure_medical_response(raw)
risk_level, risk_score, risk_factors = detect_suicide_risk(raw)
suicide_risk = {
"level": risk_level.value,
"score": risk_score,
"factors": risk_factors
}
analysis_doc = {
"identifier": patient_id,
"patient_id": patient_id,
"timestamp": datetime.utcnow(),
"summary": structured,
"suicide_risk": suicide_risk,
"raw": raw,
"data_hash": patient_hash
}
await analysis_collection.update_one(
{"identifier": patient_id},
{"$set": analysis_doc},
upsert=True
)
if risk_level in [RiskLevel.MODERATE, RiskLevel.HIGH, RiskLevel.SEVERE]:
await create_alert(patient_id, suicide_risk)
logger.info(f"✅ Stored analysis for patient {patient_id}")
except Exception as e:
logger.error(f"Error analyzing patient: {e}")
def recognize_speech(audio_data: bytes, language: str = "en-US") -> str:
recognizer = sr.Recognizer()
try:
with io.BytesIO(audio_data) as audio_file:
with sr.AudioFile(audio_file) as source:
audio = recognizer.record(source)
text = recognizer.recognize_google(audio, language=language)
return text
except sr.UnknownValueError:
logger.error("Google Speech Recognition could not understand audio")
raise HTTPException(status_code=400, detail="Could not understand audio")
except sr.RequestError as e:
logger.error(f"Could not request results from Google Speech Recognition service; {e}")
raise HTTPException(status_code=503, detail="Speech recognition service unavailable")
except Exception as e:
logger.error(f"Error in speech recognition: {e}")
raise HTTPException(status_code=500, detail="Error processing speech")
def text_to_speech(text: str, language: str = "en", slow: bool = False) -> bytes:
try:
tts = gTTS(text=text, lang=language, slow=slow)
mp3_fp = io.BytesIO()
tts.write_to_fp(mp3_fp)
mp3_fp.seek(0)
return mp3_fp.read()
except Exception as e:
logger.error(f"Error in text-to-speech conversion: {e}")
raise HTTPException(status_code=500, detail="Error generating speech")
@app.on_event("startup")
async def startup_event():
global agent, patients_collection, analysis_collection, alerts_collection
agent = TxAgent(
model_name="mims-harvard/TxAgent-T1-Llama-3.1-8B",
rag_model_name="mims-harvard/ToolRAG-T1-GTE-Qwen2-1.5B",
enable_finish=True,
enable_rag=False,
force_finish=True,
enable_checker=True,
step_rag_num=4,
seed=42
)
agent.chat_prompt = (
"You are a clinical assistant AI. Analyze the patient's data and provide clear clinical recommendations."
)
agent.init_model()
logger.info("✅ TxAgent initialized")
db = get_mongo_client()["cps_db"]
global users_collection # Add this to access users_collection for authentication
users_collection = db["users"]
patients_collection = db["patients"]
analysis_collection = db["patient_analysis_results"]
alerts_collection = db["clinical_alerts"]
logger.info("📡 Connected to MongoDB")
asyncio.create_task(analyze_all_patients())
# Protected Endpoints (add Depends(get_current_user) to all endpoints)
@app.get("/status")
async def status(current_user: dict = Depends(get_current_user)):
logger.info(f"Status endpoint accessed by {current_user['email']}")
return {
"status": "running",
"timestamp": datetime.utcnow().isoformat(),
"version": "2.6.0",
"features": ["chat", "voice-input", "voice-output", "patient-analysis", "report-upload"]
}
@app.get("/patients/analysis-results")
async def get_patient_analysis_results(
name: Optional[str] = Query(None),
current_user: dict = Depends(get_current_user)
):
logger.info(f"Fetching analysis results by {current_user['email']}")
try:
query = {}
if name:
name_regex = re.compile(name, re.IGNORECASE)
matching_patients = await patients_collection.find({"full_name": name_regex}).to_list(length=None)
patient_ids = [p["fhir_id"] for p in matching_patients if "fhir_id" in p]
if not patient_ids:
return []
query = {"patient_id": {"$in": patient_ids}}
analyses = await analysis_collection.find(query).sort("timestamp", -1).to_list(length=100)
enriched_results = []
for analysis in analyses:
patient = await patients_collection.find_one({"fhir_id": analysis.get("patient_id")})
if patient:
analysis["full_name"] = patient.get("full_name", "Unknown")
analysis["_id"] = str(analysis["_id"])
enriched_results.append(analysis)
return enriched_results
except Exception as e:
logger.error(f"Error fetching analysis results: {e}")
raise HTTPException(status_code=500, detail="Failed to retrieve analysis results")
@app.post("/chat-stream")
async def chat_stream_endpoint(
request: ChatRequest,
current_user: dict = Depends(get_current_user)
):
logger.info(f"Chat stream initiated by {current_user['email']}")
async def token_stream():
try:
conversation = [{"role": "system", "content": agent.chat_prompt}]
if request.history:
conversation.extend(request.history)
conversation.append({"role": "user", "content": request.message})
input_ids = agent.tokenizer.apply_chat_template(
conversation, add_generation_prompt=True, return_tensors="pt"
).to(agent.device)
output = agent.model.generate(
input_ids,
do_sample=True,
temperature=request.temperature,
max_new_tokens=request.max_new_tokens,
pad_token_id=agent.tokenizer.eos_token_id,
return_dict_in_generate=True
)
text = agent.tokenizer.decode(output["sequences"][0][input_ids.shape[1]:], skip_special_tokens=True)
for chunk in text.split():
yield chunk + " "
await asyncio.sleep(0.05)
except Exception as e:
logger.error(f"Streaming error: {e}")
yield f"⚠️ Error: {e}"
return StreamingResponse(token_stream(), media_type="text/plain")
@app.post("/voice/transcribe")
async def transcribe_voice(
audio: UploadFile = File(...),
language: str = Query("en-US", description="Language code for speech recognition"),
current_user: dict = Depends(get_current_user)
):
logger.info(f"Voice transcription initiated by {current_user['email']}")
try:
audio_data = await audio.read()
if not audio.filename.lower().endswith(('.wav', '.mp3', '.ogg', '.flac')):
raise HTTPException(status_code=400, detail="Unsupported audio format")
text = recognize_speech(audio_data, language)
return {"text": text}
except HTTPException:
raise
except Exception as e:
logger.error(f"Error in voice transcription: {e}")
raise HTTPException(status_code=500, detail="Error processing voice input")
@app.post("/voice/synthesize")
async def synthesize_voice(
request: VoiceOutputRequest,
current_user: dict = Depends(get_current_user)
):
logger.info(f"Voice synthesis initiated by {current_user['email']}")
try:
audio_data = text_to_speech(request.text, request.language, request.slow)
if request.return_format == "base64":
return {"audio": base64.b64encode(audio_data).decode('utf-8')}
else:
return StreamingResponse(
io.BytesIO(audio_data),
media_type="audio/mpeg",
headers={"Content-Disposition": "attachment; filename=speech.mp3"}
)
except HTTPException:
raise
except Exception as e:
logger.error(f"Error in voice synthesis: {e}")
raise HTTPException(status_code=500, detail="Error generating voice output")
@app.post("/voice/chat")
async def voice_chat_endpoint(
audio: UploadFile = File(...),
language: str = Query("en-US", description="Language code for speech recognition"),
temperature: float = Query(0.7, ge=0.1, le=1.0),
max_new_tokens: int = Query(512, ge=50, le=1024),
current_user: dict = Depends(get_current_user)
):
logger.info(f"Voice chat initiated by {current_user['email']}")
try:
audio_data = await audio.read()
user_message = recognize_speech(audio_data, language)
chat_response = agent.chat(
message=user_message,
history=[],
temperature=temperature,
max_new_tokens=max_new_tokens
)
audio_data = text_to_speech(chat_response, language.split('-')[0])
return StreamingResponse(
io.BytesIO(audio_data),
media_type="audio/mpeg",
headers={"Content-Disposition": "attachment; filename=response.mp3"}
)
except HTTPException:
raise
except Exception as e:
logger.error(f"Error in voice chat: {e}")
raise HTTPException(status_code=500, detail="Error processing voice chat")
@app.post("/analyze-report")
async def analyze_clinical_report(
file: UploadFile = File(...),
patient_id: Optional[str] = Form(None),
temperature: float = Form(0.5),
max_new_tokens: int = Form(1024),
current_user: dict = Depends(get_current_user)
):
logger.info(f"Report analysis initiated by {current_user['email']}")
try:
content_type = file.content_type
allowed_types = [
'application/pdf',
'text/plain',
'application/vnd.openxmlformats-officedocument.wordprocessingml.document'
]
if content_type not in allowed_types:
raise HTTPException(
status_code=400,
detail=f"Unsupported file type: {content_type}. Supported types: PDF, TXT, DOCX"
)
file_content = await file.read()
if content_type == 'application/pdf':
text = extract_text_from_pdf(file_content)
elif content_type == 'text/plain':
text = file_content.decode('utf-8')
elif content_type == 'application/vnd.openxmlformats-officedocument.wordprocessingml.document':
doc = Document(io.BytesIO(file_content))
text = "\n".join([para.text for para in doc.paragraphs])
else:
raise HTTPException(status_code=400, detail="Unsupported file type")
text = clean_text_response(text)
if len(text.strip()) < 50:
raise HTTPException(
status_code=400,
detail="Extracted text is too short (minimum 50 characters required)"
)
analysis = await analyze_patient_report(
patient_id=patient_id,
report_content=text,
file_type=content_type,
file_content=file_content
)
if "_id" in analysis and isinstance(analysis["_id"], ObjectId):
analysis["_id"] = str(analysis["_id"])
if "timestamp" in analysis and isinstance(analysis["timestamp"], datetime):
analysis["timestamp"] = analysis["timestamp"].isoformat()
return JSONResponse(content=jsonable_encoder({
"status": "success",
"analysis": analysis,
"patient_id": patient_id,
"file_type": content_type,
"file_size": len(file_content)
}))
except HTTPException:
raise
except Exception as e:
logger.error(f"Error in report analysis: {str(e)}")
raise HTTPException(
status_code=500,
detail=f"Failed to analyze report: {str(e)}"
)
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8000)