feat(ssa): Complete T-test end-to-end testing with 9 bug fixes - Phase 1 core 85% complete. R service: missing value auto-filter. Backend: error handling, variable matching, dynamic filename. Frontend: module activation, session isolation, error propagation. Full flow verified.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-02-19 20:57:00 +08:00
parent 8137e3cde2
commit 49b5c37cb1
86 changed files with 21207 additions and 252 deletions
--- a/backend/src/modules/iit-manager/engines/SoftRuleEngine.js
+++ b/backend/src/modules/iit-manager/engines/SoftRuleEngine.js
@@ -0,0 +1,353 @@
+/**
+ * SoftRuleEngine - 软规则质控引擎 (LLM 推理)
+ *
+ * 功能：
+ * - 调用 LLM 进行复杂的医学逻辑判断
+ * - 支持入排标准、AE 事件检测、方案偏离等场景
+ * - 返回带证据链的结构化结果
+ *
+ * 设计原则：
+ * - 智能推理：利用 LLM 处理模糊规则和复杂逻辑
+ * - 证据链：每个判断都附带推理过程和证据
+ * - 三态输出：PASS / FAIL / UNCERTAIN（需人工确认）
+ */
+import { PrismaClient } from '@prisma/client';
+import { LLMFactory } from '../../../common/llm/adapters/LLMFactory.js';
+import { logger } from '../../../common/logging/index.js';
+import { buildClinicalSlice } from '../services/PromptBuilder.js';
+const prisma = new PrismaClient();
+// ============================================================
+// SoftRuleEngine 实现
+// ============================================================
+export class SoftRuleEngine {
+    constructor(projectId, config) {
+        this.projectId = projectId;
+        this.model = config?.model || 'deepseek-v3';
+        this.timeoutMs = config?.timeoutMs || 30000;
+    }
+    /**
+     * 执行软规则检查
+     *
+     * @param recordId 记录ID
+     * @param data 记录数据
+     * @param checks 要执行的检查列表
+     * @returns 检查结果
+     */
+    async execute(recordId, data, checks) {
+        const startTime = Date.now();
+        const results = [];
+        const failedChecks = [];
+        const uncertainChecks = [];
+        logger.info('[SoftRuleEngine] Starting execution', {
+            projectId: this.projectId,
+            recordId,
+            checkCount: checks.length,
+            model: this.model,
+        });
+        // 逐个执行检查（可以改为并发，但需注意 Token 限制）
+        for (const check of checks) {
+            try {
+                const result = await this.executeCheck(recordId, data, check);
+                results.push(result);
+                if (result.status === 'FAIL') {
+                    failedChecks.push(result);
+                }
+                else if (result.status === 'UNCERTAIN') {
+                    uncertainChecks.push(result);
+                }
+            }
+            catch (error) {
+                logger.error('[SoftRuleEngine] Check execution failed', {
+                    checkId: check.id,
+                    error: error.message,
+                });
+                // 发生错误时标记为 UNCERTAIN
+                const errorResult = {
+                    checkId: check.id,
+                    checkName: check.name,
+                    status: 'UNCERTAIN',
+                    reason: `执行出错: ${error.message}`,
+                    evidence: {},
+                    confidence: 0,
+                    severity: check.severity,
+                    category: check.category,
+                };
+                results.push(errorResult);
+                uncertainChecks.push(errorResult);
+            }
+        }
+        // 计算整体状态
+        let overallStatus = 'PASS';
+        if (failedChecks.length > 0) {
+            overallStatus = 'FAIL';
+        }
+        else if (uncertainChecks.length > 0) {
+            overallStatus = 'UNCERTAIN';
+        }
+        const duration = Date.now() - startTime;
+        logger.info('[SoftRuleEngine] Execution completed', {
+            recordId,
+            overallStatus,
+            totalChecks: checks.length,
+            failed: failedChecks.length,
+            uncertain: uncertainChecks.length,
+            duration: `${duration}ms`,
+        });
+        return {
+            recordId,
+            projectId: this.projectId,
+            timestamp: new Date().toISOString(),
+            overallStatus,
+            summary: {
+                totalChecks: checks.length,
+                passed: results.filter(r => r.status === 'PASS').length,
+                failed: failedChecks.length,
+                uncertain: uncertainChecks.length,
+            },
+            results,
+            failedChecks,
+            uncertainChecks,
+        };
+    }
+    /**
+     * 执行单个检查
+     */
+    async executeCheck(recordId, data, check) {
+        const startTime = Date.now();
+        // 1. 构建 Prompt
+        const prompt = this.buildCheckPrompt(recordId, data, check);
+        // 2. 调用 LLM
+        const llmAdapter = LLMFactory.getAdapter(this.model);
+        const response = await llmAdapter.chat([
+            {
+                role: 'system',
+                content: this.getSystemPrompt(),
+            },
+            {
+                role: 'user',
+                content: prompt,
+            },
+        ]);
+        const rawResponse = response.content;
+        // 3. 解析响应
+        const parsed = this.parseResponse(rawResponse, check);
+        const duration = Date.now() - startTime;
+        logger.debug('[SoftRuleEngine] Check executed', {
+            checkId: check.id,
+            status: parsed.status,
+            confidence: parsed.confidence,
+            duration: `${duration}ms`,
+        });
+        return {
+            checkId: check.id,
+            checkName: check.name,
+            status: parsed.status,
+            reason: parsed.reason,
+            evidence: parsed.evidence,
+            confidence: parsed.confidence,
+            severity: check.severity,
+            category: check.category,
+            rawResponse,
+        };
+    }
+    /**
+     * 构建检查 Prompt
+     */
+    buildCheckPrompt(recordId, data, check) {
+        // 使用 PromptBuilder 生成临床数据切片
+        const clinicalSlice = buildClinicalSlice({
+            task: check.name,
+            criteria: [check.description || check.name],
+            patientData: data,
+            tags: check.requiredTags,
+            instruction: '请根据以下数据进行判断。',
+        });
+        // 替换 Prompt 模板中的变量
+        let userPrompt = check.promptTemplate;
+        // 替换 {{variable}} 格式的占位符
+        userPrompt = userPrompt.replace(/\{\{(\w+)\}\}/g, (_, key) => {
+            return data[key] !== undefined ? String(data[key]) : `[${key}未提供]`;
+        });
+        // 替换 {{#tag}} 格式的数据标签
+        userPrompt = userPrompt.replace(/\{\{#(\w+)\}\}/g, (_, tag) => {
+            // 根据标签筛选相关字段
+            return JSON.stringify(data, null, 2);
+        });
+        return `${clinicalSlice}\n\n---\n\n## 检查任务\n\n${userPrompt}`;
+    }
+    /**
+     * 获取系统 Prompt
+     */
+    getSystemPrompt() {
+        return `你是一个专业的临床研究数据监查员 (CRA)，负责核查受试者数据的质量和合规性。
+
+## 你的职责
+1. 仔细分析提供的临床数据
+2. 根据检查任务进行判断
+3. 给出清晰的判断结果和理由
+
+## 输出格式要求
+请严格按照以下 JSON 格式输出：
+
+\`\`\`json
+{
+  "status": "PASS" | "FAIL" | "UNCERTAIN",
+  "reason": "判断理由的详细说明",
+  "evidence": {
+    "key_field_1": "相关数据值",
+    "key_field_2": "相关数据值"
+  },
+  "confidence": 0.95
+}
+\`\`\`
+
+## 状态说明
+- **PASS**: 检查通过，数据符合要求
+- **FAIL**: 检查失败，发现问题
+- **UNCERTAIN**: 数据不足或存在歧义，需要人工确认
+
+## 置信度说明
+- 0.9-1.0: 非常确定
+- 0.7-0.9: 比较确定
+- 0.5-0.7: 有一定把握
+- <0.5: 不太确定，建议人工复核
+
+请只输出 JSON，不要有其他内容。`;
+    }
+    /**
+     * 解析 LLM 响应
+     */
+    parseResponse(rawResponse, check) {
+        try {
+            // 尝试提取 JSON
+            const jsonMatch = rawResponse.match(/```json\s*([\s\S]*?)\s*```/);
+            const jsonStr = jsonMatch ? jsonMatch[1] : rawResponse;
+            const parsed = JSON.parse(jsonStr.trim());
+            // 验证状态值
+            const validStatuses = ['PASS', 'FAIL', 'UNCERTAIN'];
+            const status = validStatuses.includes(parsed.status?.toUpperCase())
+                ? parsed.status.toUpperCase()
+                : 'UNCERTAIN';
+            return {
+                status: status,
+                reason: parsed.reason || '未提供理由',
+                evidence: parsed.evidence || {},
+                confidence: typeof parsed.confidence === 'number'
+                    ? Math.min(1, Math.max(0, parsed.confidence))
+                    : 0.5,
+            };
+        }
+        catch (error) {
+            logger.warn('[SoftRuleEngine] Failed to parse LLM response', {
+                checkId: check.id,
+                rawResponse: rawResponse.substring(0, 500),
+            });
+            // 解析失败时尝试简单匹配
+            const lowerResponse = rawResponse.toLowerCase();
+            if (lowerResponse.includes('pass') || lowerResponse.includes('通过')) {
+                return {
+                    status: 'PASS',
+                    reason: rawResponse,
+                    evidence: {},
+                    confidence: 0.6,
+                };
+            }
+            else if (lowerResponse.includes('fail') || lowerResponse.includes('失败') || lowerResponse.includes('不符合')) {
+                return {
+                    status: 'FAIL',
+                    reason: rawResponse,
+                    evidence: {},
+                    confidence: 0.6,
+                };
+            }
+            return {
+                status: 'UNCERTAIN',
+                reason: `无法解析响应: ${rawResponse.substring(0, 200)}`,
+                evidence: {},
+                confidence: 0.3,
+            };
+        }
+    }
+    /**
+     * 批量执行检查
+     *
+     * @param records 记录列表
+     * @param checks 检查列表
+     * @returns 所有记录的检查结果
+     */
+    async executeBatch(records, checks) {
+        const results = [];
+        for (const record of records) {
+            const result = await this.execute(record.recordId, record.data, checks);
+            results.push(result);
+        }
+        return results;
+    }
+}
+// ============================================================
+// 工厂函数
+// ============================================================
+/**
+ * 创建 SoftRuleEngine 实例
+ *
+ * @param projectId 项目ID
+ * @param config 可选配置
+ * @returns SoftRuleEngine 实例
+ */
+export function createSoftRuleEngine(projectId, config) {
+    return new SoftRuleEngine(projectId, config);
+}
+// ============================================================
+// 预置检查模板
+// ============================================================
+/**
+ * 入排标准检查模板
+ */
+export const INCLUSION_EXCLUSION_CHECKS = [
+    {
+        id: 'IE-001',
+        name: '年龄入组标准',
+        description: '检查受试者年龄是否符合入组标准',
+        promptTemplate: '请根据受试者数据，判断其年龄是否在研究方案规定的入组范围内。如果年龄字段缺失，请标记为 UNCERTAIN。',
+        requiredTags: ['#demographics'],
+        category: 'inclusion',
+        severity: 'critical',
+    },
+    {
+        id: 'IE-002',
+        name: '确诊时间入组标准',
+        description: '检查受试者确诊时间是否符合入组标准（通常要求确诊在一定时间内）',
+        promptTemplate: '请根据受试者的确诊日期和入组日期，判断确诊时间是否符合研究方案要求。',
+        requiredTags: ['#demographics', '#medical_history'],
+        category: 'inclusion',
+        severity: 'critical',
+    },
+];
+/**
+ * AE 事件检测模板
+ */
+export const AE_DETECTION_CHECKS = [
+    {
+        id: 'AE-001',
+        name: 'Lab 异常与 AE 一致性',
+        description: '检查实验室检查异常值是否已在 AE 表中报告',
+        promptTemplate: '请对比实验室检查数据和不良事件记录，判断是否存在未报告的实验室异常（Grade 3 及以上）。',
+        requiredTags: ['#lab', '#ae'],
+        category: 'ae_detection',
+        severity: 'critical',
+    },
+];
+/**
+ * 方案偏离检测模板
+ */
+export const PROTOCOL_DEVIATION_CHECKS = [
+    {
+        id: 'PD-001',
+        name: '访视超窗检测',
+        description: '检查访视是否在方案规定的时间窗口内',
+        promptTemplate: '请根据访视日期数据，判断各访视之间的时间间隔是否符合方案规定的访视窗口。',
+        requiredTags: ['#visits'],
+        category: 'protocol_deviation',
+        severity: 'warning',
+    },
+];