AI增强HEp-2间接免疫荧光图像分析预测系统性自身免疫风湿病特异性自身抗体的开发与时间验证
该研究开发了针对ANA阳性患者HEp-2 IIF图像的人工智能模型,用于预测特定自身抗体。在47,890例开发队列中,CMA、dsDNA、Sm、SSA和SSB模型的AUROC≥0.80,其中CMA最高(AUROC 0.95)。在前瞻性验证队列(n=21,427)中性能相似,dsDNA判别在不同检测方法间保持稳定(AUROC约0.85)。
要点
- AI模型对CMA、dsDNA、Sm、SSA和SSB的预测AUROC≥0.80,CMA最高达0.95。
- 与简化IIF模式相比,dsDNA和Sm模型更多将抗体阴性病例分类为阴性。
- RNP(AUROC 0.68)和Scl-70(AUROC 0.59)模型性能较低。
核心结论
- AI应用于ANA IIF图像可识别特定自身抗体,并常改善抗体阴性病例的分类。
- 研究结果支持其标准化ANA解读及分诊或减少确认试验的潜力。
研究设计与样本
- 开发队列纳入2016年12月12日至2022年12月31日ANA阳性(IIF滴度≥1:80)且90天内检测特定自身抗体的患者,取首次ANA检测。
- 前瞻性队列纳入2023年1月1日至2024年8月20日患者,用于时间验证,期间包含dsDNA检测方法转换。
- 共筛选512,959例患者,47,890例符合纳入标准,平均年龄55岁(SD 17.6),女性占78%。
主要发现
- CMA、dsDNA、Sm、SSA和SSB模型AUROC≥0.80,CMA最高(AUROC 0.95,准确率94.2%,特异性95.3%)。
- 与相关IIF模式相比,dsDNA和Sm模型更多将抗体阴性病例分类为阴性(dsDNA 88.2% vs 54.7%;Sm 82.6% vs 60.3%),而抗体阳性病例分类相似。
- RNP模型AUROC 0.68,Scl-70模型AUROC 0.59,性能较低。
- 前瞻性队列(n=21,427)中性能相似,dsDNA判别在不同检测方法间保持稳定(AUROC约0.85)。
临床怎么用
- AI辅助ANA IIF图像分析可能帮助标准化解读,并用于分诊或减少不必要的确认试验。
局限性
- 原文未明确报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: Antinuclear antibody (ANA) indirect immunofluorescence (IIF) is the reference standard but depends on subjective pattern interpretation. We tested whether artificial intelligence (AI) applied to ANA IIF images could identify specific autoantibodies. METHODS: We assembled a cohort of ANA-positive patients (human epithelial type 2 IIF titer ≥1:80) with anti-double-stranded DNA (dsDNA), anti-Smith (Sm) antibody, ribonucleoprotein (RNP), anti-Sjögren's disease-related antigen A (SSA), anti-Sjögren's disease-related antigen B (SSB), anti-topoisomerase I (Scl-70), or centromere antibody (CMA) testing within 90 days. The development cohort (December 12, 2016-December 31, 2022) included the first ANA test per patient. Two AI architectures were trained for each autoantibody. Performance was assessed using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy; comparisons with IIF patterns were summarized as pattern-antibody agreement. A prospective cohort (January 1, 2023-August 20, 2024), including a dsDNA assay transition, provided temporal validation. RESULTS: Among 512,959 patients, 47,890 met inclusion criteria (mean age 55 [SD 17.6], 78% female). Models for CMA, dsDNA, Sm, SSA, and SSB achieved AUROC ≥0.80 (CMA highest: AUROC 0.95, accuracy 94.2%, specificity 95.3%). Compared with associated IIF patterns, dsDNA and Sm models more often classified antibody-negative cases as negative (dsDNA 88.2% vs 54.7%; Sm 82.6% vs 60.3%) while similarly classifying antibody-positive cases. Performance for RNP (AUROC 0.68) and Scl-70 (AUROC 0.59) was lower. In the prospective cohort (n = 21,427), performance was similar; dsDNA discrimination remained stable across assays (AUROC ~0.85). CONCLUSION: AI applied to ANA IIF images identified specific autoantibodies and often improved classification of antibody-negative cases versus simplified pattern-based comparators. These findings support its potential to standardize ANA interpretation and triage or reduce confirmatory testing.
引用
- Vanderboom P, Dasari S, Sandino-Bermúdez MJ, et al. Development and Temporal Validation of AI-Enhanced Human Epithelial Type 2 Indirect Immunofluorescence Image Analysis for Specific Autoantibody Prediction in Systemic Autoimmune Rheumatic Diseases. Arthritis Rheumatol. 2026. PMID: 42663162. doi:10.1002/art.70312
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。