Research Papers 论文研究 8d ago Updated 7d ago 更新于 7天前 43

AnchorSIPS: A Synthetic Dataset and Evaluation Resource for Evidence-Supported Psychosis-Risk Symptom Measurement AnchorSIPS:用于循证精神病风险症状测量的合成数据集与评估资源

AnchorSIPS is a synthetic dataset of 10,000 structured psychosis-risk interviews modeled on the Mini-SIPS clinical interview, designed to bypass privacy and consent barriers that limit real clinical data sharing. The dataset uses a plan-then-realize generation pipeline: a hidden case sheet defines the clinical state, a deterministic planner fixes interview structure, and an LLM generates only patient utterances under validation and bounded repair to avoid inter-turn inconsistencies. Every interm 提出AnchorSIPS:首个面向精神病风险评估的合成数据集,包含10K条结构化访谈,解决临床数据隐私与共享瓶颈 采用"规划-实现"生成管道:隐藏病例表→确定性规划器固定访谈结构→LLM仅生成患者发言,避免多轮对话不一致性 访谈基于Mini-SIPS标准,覆盖24项症状问题、支持证据提取、幻觉/妄想类症状判定及最终APS诊断 七种LLM基线测试显示:模型能恢复粗粒度决策,但无法提取随访细节或引用支持性对话轮次,最终标签性能高估了实际能力

55
Hot 热度
72
Quality 质量
62
Impact 影响力

Analysis 深度分析

TL;DR

  • AnchorSIPS is a synthetic dataset of 10,000 structured psychosis-risk interviews modeled on the Mini-SIPS clinical interview, designed to bypass privacy and consent barriers that limit real clinical data sharing.
  • The dataset uses a plan-then-realize generation pipeline: a hidden case sheet defines the clinical state, a deterministic planner fixes interview structure, and an LLM generates only patient utterances under validation and bounded repair to avoid inter-turn inconsistencies.
  • Every intermediate decision (symptom endorsements, delusion/hallucination classifications, frank-psychosis exclusion, and final APS diagnosis) is anchored to supporting transcript turns, enabling evidence-grounded evaluation.
  • Seven LLM baselines recovered coarse diagnostic decisions but failed to extract follow-up evidence or cite supporting transcript turns, revealing that final-label accuracy overstates true interview competence.
  • AnchorSIPS is positioned as a research resource for evidence extraction, transcript-grounded measurement, and studying model uncertainty under partial disclosure in clinical NLP.

Why It Matters

This work addresses a critical bottleneck in AI-driven mental health research: the scarcity of shareable clinical interview data due to privacy, governance, and consent constraints. By providing a rigorously structured synthetic dataset with transcript-anchored labels, AnchorSIPS enables reproducible evaluation of evidence extraction and diagnostic reasoning—capabilities essential for any AI system intended to support psychosis-risk assessment in real clinical settings.

Technical Details

  • Dataset scale and structure: 10,000 synthetic interviews, each containing patient history, 24 symptom questions, follow-up evidence for affirmed items, symptom-class decisions (delusion-like, hallucination-like, disorganized communication), frank-psychosis exclusion checks, and a final Attenuated Psychosis Syndrome (APS) diagnosis.
  • Generation pipeline: A three-stage plan-then-realize architecture: (1) a hidden case sheet specifies the patient's clinical state, (2) a deterministic planner fixes the interview structure and all diagnostic labels, and (3) an LLM generates only patient utterances under validation and bounded repair, preventing the inter-turn inconsistencies common in multi-turn LLM dialogue.
  • Transcript grounding: Every intermediate decision is linked to specific supporting transcript turns, making the dataset suitable for evidence extraction and citation-based evaluation rather than simple label prediction.
  • Evaluation: Seven LLM baselines were tested; models performed adequately on coarse diagnostic decisions but significantly underperformed on extracting follow-up details and citing supporting evidence, demonstrating a gap between label-level and evidence-level competence.
  • Target research directions: Evidence extraction, transcript-grounded clinical measurement, and uncertainty quantification under partial disclosure.

Industry Insight

  • Synthetic clinical datasets with rigorous grounding structures will become increasingly important as privacy regulations tighten; AnchorSIPS demonstrates a viable blueprint for plan-then-realize generation that prioritizes structural fidelity over raw fluency.
  • The finding that label-level performance overstates interview competence is a cautionary signal for the mental health AI community: evaluation frameworks must demand evidence citation and intermediate reasoning, not just final diagnostic accuracy.
  • Researchers building clinical NLP systems should treat AnchorSIPS as a benchmark for evidence extraction and consider adopting its transcript-anchoring methodology to ensure their models can justify decisions with supporting text, a requirement for any system approaching real-world clinical deployment.

TL;DR

  • 提出AnchorSIPS:首个面向精神病风险评估的合成数据集,包含10K条结构化访谈,解决临床数据隐私与共享瓶颈
  • 采用"规划-实现"生成管道:隐藏病例表→确定性规划器固定访谈结构→LLM仅生成患者发言,避免多轮对话不一致性
  • 访谈基于Mini-SIPS标准,覆盖24项症状问题、支持证据提取、幻觉/妄想类症状判定及最终APS诊断
  • 七种LLM基线测试显示:模型能恢复粗粒度决策,但无法提取随访细节或引用支持性对话轮次,最终标签性能高估了实际能力

为什么值得看

本文为AI精神健康评估领域提供了首个可公开使用的结构化合成数据集,突破临床数据隐私壁垒。其"规划-实现"生成范式为医疗对话数据合成提供了可复用的技术路径,同时揭示了当前LLM在证据链追踪方面的能力短板。

技术解析

  • 数据集规模与结构:10,000条结构化精神病风险访谈,每条包含病史、24项症状问题、患者确认后的随访证据、症状分类决策(妄想样/幻觉样/紊乱沟通)、排除 frank psychosis 检查,以及最终的APS诊断标签
  • 生成架构:三层管道设计——(1) 隐藏病例表指定患者临床状态;(2) 确定性规划器固定访谈逻辑结构;(3) LLM仅生成患者发言,经验证与有界修复(bounded repair)处理
  • 锚定机制:所有中间决策均锚定到支持性对话轮次,APS诊断非独立标签,而是依赖前期确认、随访细节、症状分类及排除检查的综合结果
  • 基准测试:七种LLM基线模型评估,发现模型在粗粒度决策上表现尚可,但在细粒度证据提取和引用支持性对话方面存在显著缺陷

行业启示

  • 合成数据价值验证:在医疗敏感领域,合成数据可有效突破隐私合规瓶颈,为AI模型训练提供可共享的高质量数据源
  • 评估指标警示:仅依赖最终诊断标签会高估模型能力,需建立包含证据链完整性的细粒度评估体系
  • 技术路线参考:"规划-实现"范式可推广至其他结构化医疗对话生成场景,确保逻辑一致性的同时降低LLM幻觉风险

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Dataset 数据集 Evaluation 评测 Healthcare AI 医疗AI Research 科学研究