Refine AI review false-positive rules

2026-05-27 16:52:53 +08:00
parent 54ea338bd6
commit 16b7b927d9
2 changed files with 159 additions and 13 deletions
--- a/患者首页处理/数据可视化网页端/app/main.py
+++ b/患者首页处理/数据可视化网页端/app/main.py
@@ -2346,8 +2346,11 @@ def build_ai_prompt(record: dict[str, Any], pdf_context: dict[str, Any], privacy
 7. 手术操作名称可能因为换行被结构化解析截断；如果 PDF定位文本或局部截图中显示完整多行名称，请把完整名称放入 suggested_updates。
 8. 门急诊诊断编码只能用于 outpatient_diagnosis_code；主要诊断请修正 discharge_diagnoses 中“诊断类别=主要诊断”的行，不要把门急诊诊断编码复制到 discharge_diagnoses[].疾病编码，除非出院诊断表格对应“疾病编码”单元格本身清楚显示该编码。
 9. PDF局部截图可能因遮挡、截屏边界或缩放只显示编码/文字前半段；如果 PDF 显示值是结构化字段值的前缀，且没有相反证据，应判为 match/ok，不要写“需确认完整编码”，也不要建议把结构化字段截短。
-10. remaining_issues 只写当前文档复核人应该特别注意的内容；不要写如何修改，不要重复 suggested_updates，不要写“无需补录/无需处理/首页原貌”这类已判定无问题的说明。
-11. 不要编造 PDF 中看不见的内容，不要输出置信度。
+10. 入院途径、离院方式、入院病情等代码字段只核对代码本身；例如 PDF 显示“2（门诊）”而结构化字段为“2”就是一致，不要要求复核代码与中文标签关系。
+11. 手术操作日期只有在日历日期确实早于入院日期时才算问题；同日或晚于入院日期均为正常，不要推测月份应改为其他月份。
+12. 手术及操作编码允许带 x 和 001/002/005/006 等扩展后缀；如果原始内容显示“54.5100x ... 005”这类拆开的后缀，应建议写入完整编码“54.5100x005”，不要要求人工确认扩展码是否有效。
+13. remaining_issues 只写当前文档复核人应该特别注意的内容；不要写如何修改，不要重复 suggested_updates，不要写“无需补录/无需处理/首页原貌”这类已判定无问题的说明。
+14. 不要编造 PDF 中看不见的内容，不要输出置信度。

 必须返回这个 JSON 结构：
 {{
@@ -2560,6 +2563,128 @@ def ai_pdf_prefix_truncation_issue(value: Any) -> bool:
    return ai_text_pdf_prefix_issue(value)


+def ai_extract_leading_code(value: Any) -> str:
+    text = str(value or "").strip()
+    text = text.replace("（", "(").replace("）", ")").replace("：", ":")
+    match = re.match(r"^\s*([A-Za-z]?\d+(?:[.\-xX]\d+)*)", text)
+    return ai_compact_compare_value(match.group(1)) if match else ""
+
+
+def ai_code_label_text(value: Any) -> bool:
+    text = str(value or "")
+    return bool(re.search(r"(代码|编码|入院途径|离院方式|入院病情|[A-Za-z_][A-Za-z0-9_]*_code)", text, flags=re.IGNORECASE))
+
+
+def ai_code_label_values_match(pdf_value: Any, structured_value: Any) -> bool:
+    pdf_code = ai_extract_leading_code(pdf_value)
+    structured_code = ai_extract_leading_code(structured_value)
+    if not pdf_code or not structured_code or pdf_code != structured_code:
+        return False
+    pdf_text = str(pdf_value or "")
+    structured_text = str(structured_value or "")
+    return pdf_text != structured_text and (
+        bool(re.search(r"[（(][^）)]{1,12}[）)]", pdf_text)) or ai_code_label_text(pdf_text + structured_text)
+    )
+
+
+def ai_text_code_label_consistent_issue(text: Any) -> bool:
+    raw = str(text or "")
+    if not raw or "PDF" not in raw or not ai_code_label_text(raw):
+        return False
+    if re.search(r"(缺失|为空|未填|未填写|漏填)", raw):
+        return False
+    pdf_match = re.search(
+        r"PDF[^，,；;。]*?(?:显示|勾选|可见|值)?[^，,；;。]*?(?:为|是|=)?\s*[\"'“”‘’]?([A-Za-z]?\d+(?:[.\-xX]\d+)*)(?:[）)\"'“”‘’]|[（(][^）)]{1,12}[）)])?",
+        raw,
+        flags=re.IGNORECASE,
+    )
+    structured_match = re.search(
+        r"(?:结构化字段|结构化值|结构化|[A-Za-z_][A-Za-z0-9_]*_code|[A-Za-z_][A-Za-z0-9_]*字段值|字段值)[^，,；;。]*?(?:是否为|为|是|=)?\s*[\"'“”‘’]?([A-Za-z]?\d+(?:[.\-xX]\d+)*)",
+        raw,
+        flags=re.IGNORECASE,
+    )
+    if structured_match and not pdf_match and re.search(r"(与PDF一致|与PDF相符)", raw):
+        return True
+    if pdf_match and not structured_match and re.search(r"(与结构化一致|与PDF一致|是否与结构化一致|是否匹配|是否正确)", raw):
+        return True
+    if not pdf_match or not structured_match:
+        return False
+    return ai_compact_compare_value(pdf_match.group(1)) == ai_compact_compare_value(structured_match.group(1))
+
+
+def ai_code_label_consistent_issue(value: Any) -> bool:
+    if isinstance(value, dict):
+        pdf_value = value.get("pdf_value") or value.get("pdf") or value.get("PDF值") or value.get("图片值")
+        structured_value = (
+            value.get("structured_value")
+            or value.get("structured")
+            or value.get("field_value")
+            or value.get("结构化值")
+            or value.get("结构化字段")
+        )
+        if ai_code_label_values_match(pdf_value, structured_value) and ai_code_label_text(ai_join_text(value)):
+            return True
+        return ai_text_code_label_consistent_issue(ai_join_text(value))
+    return ai_text_code_label_consistent_issue(value)
+
+
+def ai_parse_date_token(value: str) -> date | None:
+    match = re.match(r"(\d{4})[-/年](\d{1,2})[-/月](\d{1,2})", value.strip())
+    if not match:
+        return None
+    try:
+        return date(int(match.group(1)), int(match.group(2)), int(match.group(3)))
+    except ValueError:
+        return None
+
+
+def ai_dates_in_text(text: str) -> list[tuple[date, int]]:
+    dates: list[tuple[date, int]] = []
+    for match in re.finditer(r"\d{4}[-/年]\d{1,2}[-/月]\d{1,2}", text):
+        parsed = ai_parse_date_token(match.group(0))
+        if parsed:
+            dates.append((parsed, match.start()))
+    return dates
+
+
+def ai_operation_date_not_early_issue(value: Any) -> bool:
+    text = ai_join_text(value)
+    if not text or not re.search(r"(手术操作日期|手术日期|手术时间)", text):
+        return False
+    if not re.search(r"(早于入院|入院前|同一天|是否早于)", text):
+        return False
+    if re.search(r"PDF.{0,24}\d{4}[-/年]\d{1,2}[-/月]\d{1,2}.{0,24}结构化(?:字段|值)", text):
+        return False
+    admission_match = re.search(r"入院(?:时间|日期)?[^0-9]{0,12}(\d{4}[-/年]\d{1,2}[-/月]\d{1,2})", text)
+    if not admission_match:
+        return False
+    admission_date = ai_parse_date_token(admission_match.group(1))
+    if not admission_date:
+        return False
+    operation_dates = [
+        item_date
+        for item_date, position in ai_dates_in_text(text)
+        if position < admission_match.start()
+    ]
+    if not operation_dates:
+        operation_dates = [
+            item_date
+            for item_date, position in ai_dates_in_text(text)
+            if position != admission_match.start(1)
+        ]
+    if not operation_dates:
+        return False
+    return all(item_date >= admission_date for item_date in operation_dates)
+
+
+def ai_false_positive_issue(value: Any) -> bool:
+    return (
+        ai_pdf_prefix_truncation_issue(value)
+        or ai_code_label_consistent_issue(value)
+        or ai_operation_date_not_early_issue(value)
+    )
+
+
 def ai_has_nonblank_suggested_updates(parsed: dict[str, Any]) -> bool:
    suggested_updates = parsed.get("suggested_updates")
    if not isinstance(suggested_updates, list):
@@ -2579,16 +2704,20 @@ def ai_problem_evidence(parsed: dict[str, Any]) -> list[dict[str, Any]]:
    ]


-def ai_only_pdf_prefix_truncation(parsed: dict[str, Any]) -> bool:
+def ai_only_false_positive(parsed: dict[str, Any]) -> bool:
    if ai_has_nonblank_suggested_updates(parsed):
        return False
    issues = ai_remaining_issues(parsed)
    if issues:
-        return all(ai_pdf_prefix_truncation_issue(issue) for issue in issues)
+        return all(ai_false_positive_issue(issue) for issue in issues)
    evidence = ai_problem_evidence(parsed)
    if evidence:
-        return all(ai_pdf_prefix_truncation_issue(item) for item in evidence)
-    return ai_pdf_prefix_truncation_issue(parsed.get("summary")) or ai_pdf_prefix_truncation_issue(parsed.get("decision"))
+        return all(ai_false_positive_issue(item) for item in evidence)
+    return ai_false_positive_issue(parsed.get("summary")) or ai_false_positive_issue(parsed.get("decision"))
+
+
+def ai_only_pdf_prefix_truncation(parsed: dict[str, Any]) -> bool:
+    return ai_only_false_positive(parsed)


 def normalize_ai_parsed(parsed: dict[str, Any]) -> dict[str, Any]:
@@ -2596,9 +2725,9 @@ def normalize_ai_parsed(parsed: dict[str, Any]) -> dict[str, Any]:
    remaining = normalized.get("remaining_issues")
    if isinstance(remaining, list):
        normalized["remaining_issues"] = [
-            item for item in remaining if ai_needs_review_text(item) and not ai_pdf_prefix_truncation_issue(item)
+            item for item in remaining if ai_needs_review_text(item) and not ai_false_positive_issue(item)
        ]
-    elif remaining and ai_needs_review_text(remaining) and not ai_pdf_prefix_truncation_issue(remaining):
+    elif remaining and ai_needs_review_text(remaining) and not ai_false_positive_issue(remaining):
        normalized["remaining_issues"] = [remaining]
    else:
        normalized["remaining_issues"] = []
@@ -2622,7 +2751,7 @@ def ai_has_confirmed_problem(parsed: dict[str, Any]) -> bool:
            if not isinstance(item, dict):
                continue
            result = str(item.get("result") or "").strip().lower()
-            if result in {"mismatch", "uncertain", "missing", "problem"} and not ai_pdf_prefix_truncation_issue(item):
+            if result in {"mismatch", "uncertain", "missing", "problem"} and not ai_false_positive_issue(item):
                return True

    if resolution in {"false_positive", "ok", "no_issue", "误报", "无问题"}:
@@ -2651,7 +2780,7 @@ def ai_remaining_issues(parsed: dict[str, Any]) -> list[Any]:


 def ai_has_unresolved_problem(parsed: dict[str, Any]) -> bool:
-    if ai_only_pdf_prefix_truncation(parsed):
+    if ai_only_false_positive(parsed):
        return False
    if ai_remaining_issues(parsed):
        return True
@@ -2687,20 +2816,26 @@ def ai_classification(parsed: dict[str, Any]) -> str:
    resolution = str(parsed.get("issue_resolution") or "").strip().lower()
    ok_values = {"ok", "pass", "passed", "no_issue", "no issue", "无问题", "通过", "已通过"}
    problem_values = {"problem", "not_ok", "not ok", "confirm", "不ok", "不通过", "需确认", "待确认", "需复核"}
-    if ai_only_pdf_prefix_truncation(parsed):
+    if ai_only_false_positive(parsed):
        return AI_OK_STATUS
    if classification in ok_values or decision in ok_values or resolution in {"false_positive", "ok", "no_issue", "误报", "无问题", "通过"}:
        return AI_OK_STATUS
    if classification in problem_values or decision in problem_values or resolution in {"confirmed_problem", "uncertain", "problem", "待确认", "已证实"}:
        if not ai_remaining_issues(parsed) and ai_has_nonblank_suggested_updates(parsed):
            return AI_OK_STATUS
+        if (
+            not ai_remaining_issues(parsed)
+            and not any(not ai_false_positive_issue(item) for item in ai_problem_evidence(parsed))
+            and not ai_has_unresolved_problem(parsed)
+        ):
+            return AI_OK_STATUS
        return AI_PROBLEM_STATUS
    if ai_remaining_issues(parsed) or ai_has_unresolved_problem(parsed):
        return AI_PROBLEM_STATUS
    suggested_updates = parsed.get("suggested_updates")
    if isinstance(suggested_updates, list) and suggested_updates:
        return AI_OK_STATUS
-    if any(not ai_pdf_prefix_truncation_issue(item) for item in ai_problem_evidence(parsed)):
+    if any(not ai_false_positive_issue(item) for item in ai_problem_evidence(parsed)):
        return AI_PROBLEM_STATUS
    return AI_OK_STATUS

@@ -2724,6 +2859,15 @@ def blank_ai_value(value: Any) -> bool:
    return any(marker in text for marker in ("空白", "未填写", "无内容", "可见但为空", "编码栏可见但为空", "null"))


+def normalize_operation_code(value: Any) -> str:
+    text = str(value or "").strip()
+    text = re.sub(r"\s+", " ", text)
+    match = re.match(r"^([A-Za-z]?\d{1,3}\.\d{1,4}x?)\s+(\d{3})$", text, flags=re.IGNORECASE)
+    if match:
+        return f"{match.group(1)}{match.group(2)}"
+    return text
+
+
 def suggested_update_value(item: dict[str, Any]) -> Any:
    for key in ("value", "new", "new_value", "pdf_value", "suggested_value"):
        if key in item:
@@ -2829,6 +2973,8 @@ def ai_suggested_updates(result: dict[str, Any], before: dict[str, Any]) -> tupl
            compare_old = rows[index].get(key)
            if ai_outpatient_code_leak(path, value, before, compare_old):
                continue
+            if field == "operations" and key == "手术操作编码":
+                value = normalize_operation_code(value)
            if ai_pdf_value_is_structured_prefix(value, compare_old):
                continue
            if comparable(compare_old) == comparable(value):