flowchart LR
    A[患者出院请求] --> B[Agent接收并解析]
    B --> C{资格检查}
    C -->|通过| D[医疗必要性分析]
    C -->|拒绝| E[返回拒绝+原因]
    D --> F[工具调用: 查询指南/历史/费用]
    F --> G[综合评分]
    G --> H{是否批准}
    H -->|是| I[生成批准通知]
    H -->|否| J[标记拒绝+可申诉提示]

import json
from langchain.agents import Tool, AgentExecutor, create_react_agent
from langchain.memory import ConversationSummaryBufferMemory
from langchain.chat_models import ChatOpenAI
from langchain.prompts import PromptTemplate
import pandas as pd
from scipy.stats import chi2_contingency

# 模拟数据加载
guidelines = json.load(open('guidelines.json'))  # [{diagnosis, rehab_type, max_days}] 
patient_history = json.load(open('patient_history.json'))  # {patient_id: [claims]}
fees = json.load(open('fees.json'))  # {region: per_diem}

# 工具1：查询指南
def lookup_guidelines(diagnosis: str, rehab_type: str) -> dict:
    for g in guidelines:
        if g['diagnosis'] == diagnosis and g['rehab_type'] == rehab_type:
            return {"status": "success", "data": g['max_days'], "confidence": 0.9}
    return {"status": "fail", "data": "No exact match", "confidence": 0.3}

# 工具2：患者历史
def get_patient_history(patient_id: str) -> dict:
    hist = patient_history.get(patient_id, [])
    return {"status": "success", "data": hist, "confidence": 0.8}

# 工具3：费用计算
def estimate_cost(region: str, days: int) -> dict:
    per_diem = fees.get(region, 500)
    return {"status": "success", "data": per_diem * days, "confidence": 0.85}

# 工具4：先例搜索（模拟）
def search_precedent(age: int, diagnosis: str, memory) -> dict:
    # 实际应遍历记忆中的案例，这里简化
    return {"status": "success", "data": "similar case: approved (default)", "confidence": 0.5}

tools = [
    Tool(name="GuidelinesLookup", func=lookup_guidelines, description="Lookup max rehab days for diagnosis and rehab type"),
    Tool(name="PatientHistory", func=get_patient_history, description="Get patient's past claims history"),
    Tool(name="CostEstimator", func=estimate_cost, description="Estimate total cost for given region and days"),
    Tool(name="PrecedentSearch", func=lambda x: search_precedent(0, "", None), description="Search memory for similar case decisions")
]

# LLM
llm = ChatOpenAI(model="gpt-4-1106-preview", temperature=0)

# 记忆
memory = ConversationSummaryBufferMemory(llm=llm, max_token_limit=2048, return_messages=True)

# 提示模板
prompt = PromptTemplate.from_template(
    """You are a Medicare claims reviewer. Follow this plan step by step:
1. Extract patient info from input: age, diagnosis, rehab_type, region, patient_id.
2. Check guidelines for allowed days. If no exact match, use closest or ask user.
3. Check patient history for any recent denials or violations.
4. Estimate cost and compare with benefit limit ($10k).
5. Search memory for similar precedents.
6. Output final decision: APPROVED or DENIED with reason.

Use tools when needed. Output final decision in JSON format: {{"decision": "", "reason": "", "confidence": 0-1}}
User request: {input}
"""
)

agent = create_react_agent(llm=llm, tools=tools, prompt=prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, memory=memory, verbose=True, handle_parsing_errors=True)

# 示例请求
request = {
    "patient_id": "P1001",
    "age": 78,
    "diagnosis": "hip fracture",
    "rehab_type": "skilled nursing",
    "region": "NY",
    "requested_days": 20
}
result = agent_executor.invoke({"input": str(request)})
print(result['output'])

{
  "decision": "DENIED",
  "reason": "Guidelines for hip fracture allow max 14 days; requested 20 days exceeds limit. Cost estimate $10,000 exactly at cap, but no medical necessity exception documented.",
  "confidence": 0.85
}

def fairness_check(decisions_history: list) -> dict:
    df = pd.DataFrame(decisions_history)
    # 假设有 'age_group' 和 'decision' 列
    contingency = pd.crosstab(df['age_group'], df['decision'])
    if contingency.shape[1] == 2:  # approved/denied
        chi2, p, dof, expected = chi2_contingency(contingency)
        return {"chi2": chi2, "p_value": p, "significant": p < 0.05}
    return {}