AI 辅助罕见病诊断：波士顿儿童医院用 OpenAI 技术突破诊断瓶颈

"""
Rare Disease Phenotype Matcher — Prototype
依赖: pip install openai
运行前设置: export OPENAI_API_KEY="sk-..."
"""

import json
from openai import OpenAI

client = OpenAI()

# 模拟的罕见病知识库条目（实际应接入 OMIM / Orphanet）
RARE_DISEASE_DB = [
    {
        "name": "成骨不全症 (Osteogenesis Imperfecta)",
        "omim_id": "166200",
        "key_phenotypes": ["反复骨折", "蓝色巩膜", "听力下降", "短身材", "韧带松弛"],
        "gene": "COL1A1 / COL1A2"
    },
    {
        "name": "Dravet 综合征",
        "omim_id": "607208",
        "key_phenotypes": ["婴儿期热性惊厥", "难治性癫痫", "发育迟缓", "共济失调", "行为异常"],
        "gene": "SCN1A"
    },
    {
        "name": "Pompe 病 (糖原贮积病 II 型)",
        "omim_id": "232300",
        "key_phenotypes": ["肌张力低下", "心肌肥厚", "呼吸功能不全", "运动发育迟缓", "肝肿大"],
        "gene": "GAA"
    },
]

def match_phenotypes(patient_description: str, top_k: int = 3) -> list:
    """将患者表型描述与知识库进行语义匹配，返回 top_k 候选疾病。"""

    db_text = json.dumps(RARE_DISEASE_DB, ensure_ascii=False)

    system_prompt = """你是一名罕见病诊断辅助系统。你的任务是：
1. 从患者描述中提取所有临床表型关键词
2. 将这些表型与提供的罕见病知识库进行语义匹配
3. 返回最相关的疾病，按匹配度排序
4. 对每个候选疾病，列出匹配的表型和未匹配的表型

输出格式为 JSON 数组，每个元素包含:
- disease: 疾病名称
- omim_id: OMIM 编号
- matched_phenotypes: 匹配到的表型列表
- unmatched_phenotypes: 知识库中有但患者未提及的表型
- confidence: 匹配置信度 (high/medium/low)
- reasoning: 简短推理过程"""

    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": f"患者描述:\n{patient_description}\n\n罕见病知识库:\n{db_text}"},
        ],
        response_format={"type": "json_object"},
        temperature=0.2,  # 低温度，减少幻觉
    )

    result = json.loads(response.choices[0].message.content)
    return result.get("matches", result.get("results", []))[:top_k]


# --- 运行示例 ---
patient_note = """
3岁男童，自婴儿期起肌张力明显低下，运动发育里程碑全面延迟，
6个月时仍不能抬头。心脏超声示心肌肥厚，反复呼吸道感染，
近期出现肝肿大。家族史无特殊。
"""

matches = match_phenotypes(patient_note)
for m in matches:
    print(f"【{m['disease']}】(OMIM: {m['omim_id']})")
    print(f"  匹配表型: {m['matched_phenotypes']}")
    print(f"  未匹配表型: {m['unmatched_phenotypes']}")
    print(f"  置信度: {m['confidence']}")
    print(f"  推理: {m['reasoning']}")
    print()

【Pompe 病 (糖原贮积病 II 型)】(OMIM: 232300)
  匹配表型: ["肌张力低下", "心肌肥厚", "呼吸功能不全", "运动发育迟缓", "肝肿大"]
  未匹配表型: []
  置信度: high
  推理: 患者五个核心表型全部匹配，且表型组合高度特异...

【Dravet 综合征】(OMIM: 607208)
  匹配表型: ["发育迟缓"]
  ...

患者表型 → 向量嵌入 → 知识库向量检索 (top 20) → LLM 精排与推理 → 医生审阅