import sys import os import pandas as pd import pdfplumber import json import gradio as gr from typing import List, Optional from concurrent.futures import ThreadPoolExecutor, as_completed import hashlib import shutil import time from functools import lru_cache from threading import Thread import re # Environment setup current_dir = os.path.dirname(os.path.abspath(__file__)) src_path = os.path.abspath(os.path.join(current_dir, "src")) sys.path.insert(0, src_path) # Cache directories base_dir = "/data" os.makedirs(base_dir, exist_ok=True) model_cache_dir = os.path.join(base_dir, "txagent_models") tool_cache_dir = os.path.join(base_dir, "tool_cache") file_cache_dir = os.path.join(base_dir, "cache") report_dir = os.path.join(base_dir, "reports") os.makedirs(model_cache_dir, exist_ok=True) os.makedirs(tool_cache_dir, exist_ok=True) os.makedirs(file_cache_dir, exist_ok=True) os.makedirs(report_dir, exist_ok=True) os.environ.update({ "TRANSFORMERS_CACHE": model_cache_dir, "HF_HOME": model_cache_dir, "TOKENIZERS_PARALLELISM": "false", "CUDA_LAUNCH_BLOCKING": "1" }) from txagent.txagent import TxAgent MEDICAL_KEYWORDS = { 'diagnosis', 'assessment', 'plan', 'results', 'medications', 'allergies', 'summary', 'impression', 'findings', 'recommendations' } def sanitize_utf8(text: str) -> str: return text.encode("utf-8", "ignore").decode("utf-8") def file_hash(path: str) -> str: with open(path, "rb") as f: return hashlib.md5(f.read()).hexdigest() def extract_priority_pages(file_path: str, max_pages: int = 20) -> str: try: text_chunks = [] with pdfplumber.open(file_path) as pdf: for i, page in enumerate(pdf.pages[:3]): text_chunks.append(f"=== Page {i+1} ===\n{(page.extract_text() or '').strip()}") for i, page in enumerate(pdf.pages[3:max_pages], start=4): page_text = page.extract_text() or "" if any(re.search(rf'\\b{kw}\\b', page_text.lower()) for kw in MEDICAL_KEYWORDS): text_chunks.append(f"=== Page {i} ===\n{page_text.strip()}") return "\n\n".join(text_chunks) except Exception as e: return f"PDF processing error: {str(e)}" def convert_file_to_json(file_path: str, file_type: str) -> str: try: h = file_hash(file_path) cache_path = os.path.join(file_cache_dir, f"{h}.json") if os.path.exists(cache_path): return open(cache_path, "r", encoding="utf-8").read() if file_type == "pdf": text = extract_priority_pages(file_path) result = json.dumps({"filename": os.path.basename(file_path), "content": text, "status": "initial"}) Thread(target=full_pdf_processing, args=(file_path, h)).start() elif file_type == "csv": df = pd.read_csv(file_path, encoding_errors="replace", header=None, dtype=str, skip_blank_lines=False, on_bad_lines="skip") content = df.fillna("").astype(str).values.tolist() result = json.dumps({"filename": os.path.basename(file_path), "rows": content}) elif file_type in ["xls", "xlsx"]: try: df = pd.read_excel(file_path, engine="openpyxl", header=None, dtype=str) except: df = pd.read_excel(file_path, engine="xlrd", header=None, dtype=str) content = df.fillna("").astype(str).values.tolist() result = json.dumps({"filename": os.path.basename(file_path), "rows": content}) else: return json.dumps({"error": f"Unsupported file type: {file_type}"}) with open(cache_path, "w", encoding="utf-8") as f: f.write(result) return result except Exception as e: return json.dumps({"error": f"Error processing {os.path.basename(file_path)}: {str(e)}"}) def full_pdf_processing(file_path: str, file_hash: str): try: cache_path = os.path.join(file_cache_dir, f"{file_hash}_full.json") if os.path.exists(cache_path): return with pdfplumber.open(file_path) as pdf: full_text = "\n".join([f"=== Page {i+1} ===\n{(page.extract_text() or '').strip()}" for i, page in enumerate(pdf.pages)]) result = json.dumps({"filename": os.path.basename(file_path), "content": full_text, "status": "complete"}) with open(cache_path, "w", encoding="utf-8") as f: f.write(result) with open(os.path.join(report_dir, f"{file_hash}_report.txt"), "w", encoding="utf-8") as out: out.write(full_text) except Exception as e: print(f"Background processing failed: {str(e)}") def init_agent(): default_tool_path = os.path.abspath("data/new_tool.json") target_tool_path = os.path.join(tool_cache_dir, "new_tool.json") if not os.path.exists(target_tool_path): shutil.copy(default_tool_path, target_tool_path) agent = TxAgent( model_name="mims-harvard/TxAgent-T1-Llama-3.1-8B", rag_model_name="mims-harvard/ToolRAG-T1-GTE-Qwen2-1.5B", tool_files_dict={"new_tool": target_tool_path}, force_finish=True, enable_checker=True, step_rag_num=8, seed=100, additional_default_tools=[] ) agent.init_model() return agent def create_ui(agent: TxAgent): with gr.Blocks(theme=gr.themes.Soft()) as demo: gr.Markdown("

🩺 Clinical Oversight Assistant

") gr.Markdown("

Identify potential oversights in patient care

") chatbot = gr.Chatbot(label="Analysis", height=600, type="messages") file_upload = gr.File( label="Upload Medical Records", file_types=[".pdf", ".csv", ".xls", ".xlsx"], file_count="multiple" ) msg_input = gr.Textbox(placeholder="Ask about potential oversights...", show_label=False) send_btn = gr.Button("Analyze", variant="primary") conversation_state = gr.State([]) download_output = gr.File(label="Download Full Report (after tools finish)") def analyze_potential_oversights(message: str, history: list, conversation: list, files: list): start_time = time.time() try: history.append({"role": "user", "content": message}) history.append({"role": "assistant", "content": "Analyzing records for potential oversights..."}) yield history, None extracted_data = "" file_hash_value = "" if files: with ThreadPoolExecutor(max_workers=4) as executor: futures = [executor.submit(convert_file_to_json, f.name, f.name.split(".")[-1].lower()) for f in files if hasattr(f, 'name')] results = [sanitize_utf8(f.result()) for f in as_completed(futures)] extracted_data = "\n".join(results) file_hash_value = file_hash(files[0].name) analysis_prompt = f"""Review these medical records and identify EXACTLY what might have been missed: 1. List potential missed diagnoses 2. Flag any medication conflicts 3. Note incomplete assessments 4. Highlight abnormal results needing follow-up Medical Records:\n{extracted_data[:15000]} ### Potential Oversights:\n""" response = [] for chunk in agent.run_gradio_chat( message=analysis_prompt, history=[], temperature=0.2, max_new_tokens=1024, max_token=4096, call_agent=False, conversation=conversation ): if isinstance(chunk, str): response.append(chunk) elif isinstance(chunk, list): response.extend([c.content for c in chunk if hasattr(c, 'content')]) if len(response) % 3 == 0: history[-1] = {"role": "assistant", "content": "".join(response).strip()} yield history, None final_output = "".join(response).strip() if not final_output: final_output = "No clear oversights identified. Recommend comprehensive review." if not final_output.startswith(("1.", "-", "*", "#")): final_output = "• " + final_output.replace("\n", "\n• ") history[-1] = {"role": "assistant", "content": final_output} report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt") return history, report_path if os.path.exists(report_path) else None except Exception as e: history.append({"role": "assistant", "content": f"❌ Analysis failed: {str(e)}"}) return history, None inputs = [msg_input, chatbot, conversation_state, file_upload] outputs = [chatbot, download_output] send_btn.click(analyze_potential_oversights, inputs=inputs, outputs=outputs) msg_input.submit(analyze_potential_oversights, inputs=inputs, outputs=outputs) gr.Examples([ ["What might have been missed in this patient's treatment?"], ["Are there any medication conflicts in these records?"], ["What abnormal results require follow-up?"] ], inputs=msg_input) return demo if __name__ == "__main__": print("Initializing medical analysis agent...") agent = init_agent() print("Launching interface...") demo = create_ui(agent) demo.queue().launch( server_name="0.0.0.0", server_port=7860, show_error=True, share=True )