""" ============================================================================= 第 4 讲 · 工具模块：虚构的临床试验与病人 ============================================================================= 对应讲义：lectures/lecture04_accountable_ai.md 第 0.4 节（本讲的玩具世界） ----------------------------------------------------------------------------- 【这个模块做什么？】 ----------------------------------------------------------------------------- 临床试验匹配：给一个病人，判断他是否符合某个临床试验的入组条件（以及在几千个试验里找出他可能符合的那些）。 这个模块生成一个可以精确评估的玩具版本： · 一个很小的**医学本体**（ontology）：概念之间的上下位关系，例如"局灶性癫痫"是一种"癫痫"。 真实系统用 SNOMED CT 这样有几十万个概念的标准本体。 · **试验**：每个试验有一个目标疾病和若干条入组 / 排除条件（年龄、诊断、ECOG 体能状态、肌酐、 HbA1c、怀孕、HIV、心血管疾病、既往治疗……）。每条条件既有**自然语言**的写法（给 LLM 读）， 也有对应的**形式化公式**（给求解器用，见 logic.py）。 · **病人**：结构化的事实（其中一部分是**缺失**的：病历里没写），以及几种不同措辞的**病历文字** （同一个病人，换种说法 —— 用来检验"同样的事实是否得到同样的决定"）。 · **决策政策**（policy）：遇到缺失值、遇到含糊的概念时怎么办。本模块采用的政策是： P1 缺失的值视为"未知"：只要存在一种可能的取值让病人符合条件，就判为"可能合格"（进入下一步筛查）， 并且把这些取值作为**明确的假设**列出来； P2 "控制良好的高血压"**不算**"有临床意义的心血管疾病"（本体里它不在心血管疾病之下）。 标准答案 = 在这套政策下，求解器给出的判定。 【边界】这是一个极度简化的世界：真实的入组条件有几十条，涉及时间（"过去 6 个月内"）、剂量、 基因检测结果、多种疾病之间的复杂逻辑；真实的病历是几十页的自由文本，充满否定、不确定和矛盾。 """ import random from dataclasses import dataclass from logic import solve # ============================================================================= # 第 1 部分：本体 # ============================================================================= # 子概念 → 父概念。"has dx:X" 为真，当且仅当病人有某个属于 X 的具体诊断。 PARENT = { "focal epilepsy": "epilepsy", "generalized epilepsy": "epilepsy", "epilepsy": "neurological disorder", "migraine": "neurological disorder", "type 1 diabetes": "diabetes", "type 2 diabetes": "diabetes", "diabetes": "metabolic disorder", "breast cancer": "cancer", "lung cancer": "cancer", "colon cancer": "cancer", "heart failure": "cardiovascular disease", "coronary artery disease": "cardiovascular disease", "uncontrolled hypertension": "cardiovascular disease", "controlled hypertension": "hypertension", "chronic kidney disease": "kidney disease", "asthma": "respiratory disease", "COPD": "respiratory disease", } LEAF_DX = ["focal epilepsy", "generalized epilepsy", "migraine", "type 1 diabetes", "type 2 diabetes", "breast cancer", "lung cancer", "colon cancer", "heart failure", "coronary artery disease", "uncontrolled hypertension", "controlled hypertension", "chronic kidney disease", "asthma", "COPD"] CONCEPTS = sorted(set(PARENT) | set(PARENT.values())) TARGETS = ["epilepsy", "focal epilepsy", "migraine", "diabetes", "type 2 diabetes", "breast cancer", "lung cancer", "cancer", "asthma", "COPD", "chronic kidney disease", "heart failure"] TREATMENT_OF = {"epilepsy": "anticonvulsant", "diabetes": "insulin", "cancer": "chemotherapy", "respiratory disease": "inhaled steroid", "cardiovascular disease": "beta blocker", "migraine": "triptan", "kidney disease": "dialysis"} TREATMENTS = sorted(set(TREATMENT_OF.values()) | {"metformin"}) def ancestors(concept): """ 一个概念和它的所有祖先（包括它自己）。 【例子】ancestors("focal epilepsy") → {"focal epilepsy", "epilepsy", "neurological disorder"} 【智能体联系】SatIR 用 SNOMED CT 的上下位关系做同样的事：试验要求"癫痫"，有"局灶性癫痫"的病人也算。 """ out = {concept} while concept in PARENT: concept = PARENT[concept] out.add(concept) return out # ============================================================================= # 第 2 部分：试验 # ============================================================================= @dataclass class Criterion: """ 一条入组或排除条件。 【字段】 kind : "inclusion" 或 "exclusion" text : 自然语言的写法（给 LLM 读） formula : 形式化的公式（给求解器用）。排除条件的公式描述的是"被排除的情况"， 组合进试验公式时会取"非"。 cross : True = 这条条件涉及多个变量之间的逻辑（例如"ECOG 0-1，或者 70 岁以下 ECOG 2"）， 在 SatIR 的数据库投影里只能被放宽（04 号文件） """ kind: str text: str formula: tuple cross: bool = False @dataclass class Trial: """一个临床试验：编号、目标疾病、条件列表。""" tid: str target: str criteria: list def formula(self): """整个试验的公式：所有入组条件都满足，且所有排除条件都不满足。""" parts = [] for c in self.criteria: parts.append(("label", c.text, c.formula if c.kind == "inclusion" else ("not", c.formula))) return ("and", parts) def text(self): """试验的自然语言描述（给 LLM 读）。""" inc = [c.text for c in self.criteria if c.kind == "inclusion"] exc = [c.text for c in self.criteria if c.kind == "exclusion"] return (f"Trial {self.tid} for {self.target}. Inclusion criteria: " + "; ".join(inc) + ". Exclusion criteria: " + ("; ".join(exc) if exc else "none") + ".") def make_trial(tid, rng): """ 随机生成一个试验。 【思路】每个试验都有目标疾病和年龄范围；再按目标疾病的类型加上常见的条件： 癌症试验：ECOG（有时是"ECOG 0-1，或 70 岁以下 ECOG 2"这种跨变量的条件）、肌酐； 糖尿病试验：HbA1c 的范围、"未用过胰岛素"； 其他：随机加上肌酐、ECOG 等。排除条件从怀孕、HIV、心血管疾病、既往治疗里随机选。 """ target = rng.choice(TARGETS) lo, hi = rng.choice([(18, 65), (18, 75), (40, 80), (18, 85), (50, 85)]) crit = [Criterion("inclusion", f"Age {lo} to {hi} years", ("and", [("ge", "age", lo), ("le", "age", hi)])), Criterion("inclusion", f"Diagnosis of {target}", ("bool", f"dx:{target}"))] family = ancestors(target) if "cancer" in family: if rng.random() < 0.4: crit.append(Criterion("inclusion", "ECOG performance status 0-1, or 2 if younger than 70", ("or", [("le", "ecog", 1), ("and", [("le", "ecog", 2), ("le", "age", 69)])]), cross=True)) else: k = rng.choice([1, 2]) crit.append(Criterion("inclusion", f"ECOG performance status 0-{k}", ("le", "ecog", k))) crit.append(Criterion("inclusion", "Serum creatinine at most 1.5 mg/dL", ("le", "creatinine", 15))) elif "diabetes" in family: a, b = rng.choice([(70, 100), (75, 110), (65, 95)]) crit.append(Criterion("inclusion", f"HbA1c between {a / 10:.1f}% and {b / 10:.1f}%", ("and", [("ge", "hba1c", a), ("le", "hba1c", b)]))) if rng.random() < 0.5: crit.append(Criterion("exclusion", "Prior treatment with insulin", ("bool", "prior:insulin"))) else: if rng.random() < 0.5: c = rng.choice([13, 15, 20]) crit.append(Criterion("inclusion", f"Serum creatinine at most {c / 10:.1f} mg/dL", ("le", "creatinine", c))) if rng.random() < 0.3: crit.append(Criterion("inclusion", "ECOG performance status 0-2", ("le", "ecog", 2))) if rng.random() < 0.6: crit.append(Criterion("exclusion", "Pregnancy", ("bool", "pregnant"))) if rng.random() < 0.5: crit.append(Criterion("exclusion", "HIV infection", ("bool", "hiv"))) if rng.random() < 0.5: crit.append(Criterion("exclusion", "Clinically significant cardiovascular disease", ("bool", "dx:cardiovascular disease"))) if rng.random() < 0.3: t = rng.choice(TREATMENTS) if not any(c.text == f"Prior treatment with {t}" for c in crit): crit.append(Criterion("exclusion", f"Prior treatment with {t}", ("bool", f"prior:{t}"))) return Trial(tid, target, crit) # ============================================================================= # 第 3 部分：病人 # ============================================================================= @dataclass class Patient: """ 一个病人。 【字段】 pid : 编号 age, female, dx（具体诊断的列表）, prior（既往治疗的列表） labs : dict："ecog"、"creatinine"（×10 的整数，15 表示 1.5 mg/dL）、"hba1c"（×10）；缺失为 None pregnant : True / False / None（未知）；男性固定为 False hiv : True / False / None（未知） """ pid: str age: int female: bool dx: list prior: list labs: dict pregnant: object hiv: object def facts(self): """ 病人的结构化事实（求解器的输入）：变量名 → 值；未知的值为 None。 【思路】诊断按本体展开："dx:epilepsy" 为真，当且仅当病人有某种属于癫痫的具体诊断。 诊断列表视为完整的（没写的诊断就是没有）；化验、怀孕、HIV 则可能缺失。 """ f = {"age": self.age, "female": self.female, "pregnant": self.pregnant, "hiv": self.hiv} f.update(self.labs) have = set().union(*[ancestors(d) for d in self.dx]) if self.dx else set() for c in CONCEPTS: f[f"dx:{c}"] = c in have for t in TREATMENTS: f[f"prior:{t}"] = t in self.prior return f def note(self, variant=0): """ 病人的病历文字。不同的 variant 是同一组事实的不同写法（顺序、措辞、缺失值怎么表达）。 【参数】variant：0、1、2…… 【返回】str """ rng = random.Random(f"note|{self.pid}|{variant}") sex = "woman" if self.female else "man" parts = [rng.choice([f"{self.age}-year-old {sex}", f"{sex.capitalize()}, aged {self.age}", f"A {sex} of {self.age} years"]) + " with " + " and ".join(self.dx) + "."] labs = self.labs for key, name, fmt in [("ecog", "ECOG performance status", lambda v: str(v)), ("creatinine", "Serum creatinine", lambda v: f"{v / 10:.1f} mg/dL"), ("hba1c", "HbA1c", lambda v: f"{v / 10:.1f}%")]: v = labs.get(key) if v is not None: parts.append(f"{name} {fmt(v)}.") elif key != "hba1c" or any("diabetes" in d for d in self.dx): short = {"ecog": "ECOG score", "creatinine": "creatinine value", "hba1c": "HbA1c value"}[key] parts.append(rng.choice([f"{name} not reported.", f"No recent {short} on file.", ""])) if self.female: if self.pregnant is None: parts.append(rng.choice(["Pregnancy status not documented.", ""])) else: parts.append("Pregnant." if self.pregnant else rng.choice(["Not pregnant.", "Negative pregnancy test."])) if self.hiv is None: parts.append(rng.choice(["HIV status unknown.", "No HIV test on record.", ""])) else: parts.append("HIV positive." if self.hiv else rng.choice(["HIV negative.", "HIV test negative."])) parts.append("Previously treated with " + ", ".join(self.prior) + "." if self.prior else rng.choice(["No prior treatments.", "Treatment-naive."])) head, rest = parts[0], [p for p in parts[1:] if p] rng.shuffle(rest) return " ".join([head] + rest) def make_patient(pid, rng, target=None): """ 随机生成一个病人。 【参数】target：给定时，病人一定有一个属于这个概念的诊断（用来生成"和某个试验相关"的病人） """ leaves = [d for d in LEAF_DX if target and target in ancestors(d)] dx = [rng.choice(leaves)] if leaves else [rng.choice(LEAF_DX)] if rng.random() < 0.4: extra = rng.choice(LEAF_DX) if extra not in dx: dx.append(extra) if rng.random() < 0.2 and "controlled hypertension" not in dx: dx.append("controlled hypertension") # 很常见的合并症 —— 也是最容易引起歧义的一个 female = rng.random() < 0.5 age = rng.randint(14, 88) labs = {"ecog": rng.choice([0, 1, 1, 2, 2, 3]) if rng.random() < 0.7 else None, "creatinine": rng.randint(6, 22) if rng.random() < 0.6 else None, "hba1c": (rng.randint(55, 125) if rng.random() < 0.75 else None) if any("diabetes" in d for d in dx) else None} pregnant = (rng.random() < 0.08 if rng.random() < 0.7 else None) if female and age <= 45 else False hiv = (rng.random() < 0.05) if rng.random() < 0.6 else None prior = sorted({TREATMENT_OF[a] for d in dx for a in ancestors(d) if a in TREATMENT_OF and rng.random() < 0.5}) if any("diabetes" in d for d in dx) and rng.random() < 0.5: prior = sorted(set(prior) | {"metformin"}) return Patient(pid, age, female, dx, prior, labs, pregnant, hiv) TYPES = {"age": "int", "ecog": "int", "creatinine": "int", "hba1c": "int", "female": "bool", "pregnant": "bool", "hiv": "bool"} TYPES.update({f"dx:{c}": "bool" for c in CONCEPTS}) TYPES.update({f"prior:{t}": "bool" for t in TREATMENTS}) def gold_decision(trial, patient): """ 标准答案：在政策 P1、P2 下，求解器的判定。 【返回】(是否可能合格, 需要的假设) """ return solve(trial.formula(), patient.facts(), TYPES) def make_cases(n=300, seed=0): """ 生成 n 个"试验—病人"对，病人都有试验的目标疾病（相关的配对，才值得判断是否合格）。 【返回】list[(Trial, Patient)] """ rng = random.Random(seed) cases = [] for i in range(n): t = make_trial(f"T{i:04d}", rng) p = make_patient(f"P{i:04d}", rng, target=t.target) cases.append((t, p)) return cases if __name__ == "__main__": cases = make_cases(5) t, p = cases[1] print(t.text()) print("\n病历（两种写法）：") print(" " + p.note(0)) print(" " + p.note(1)) ok, assumptions = gold_decision(t, p) print(f"\n标准判定：{'可能合格' if ok else '不合格'}；需要的假设：{assumptions}") elig = sum(gold_decision(t, p)[0] for t, p in make_cases(300)) print(f"300 个相关的配对里，可能合格的：{elig}")