编辑推荐：
肺结节检出数量正在稳步增加。尽管部分结节属低危或高危，但大多数为性质未定的不确定肺结节（Indeterminate Pulmonary Nodules, IPNs）。在不确定结节中，绝大多数为良性。对不确定肺结节进行分层具有挑战性。已验证的风险计算模型因操作繁
肺结节检出数量正在稳步增加。尽管部分结节属低危或高危，但大多数为性质未定的不确定肺结节（Indeterminate Pulmonary Nodules, IPNs）。在不确定结节中，绝大多数为良性。对不确定肺结节进行分层具有挑战性。已验证的风险计算模型因操作繁琐而未能在临床常规应用。多数临床医师依赖临床直觉判断，该策略常导致不必要的活检操作，不仅费用高昂、造成患者心理压力，还存在医源性损伤风险。目前已有多项商业化肺癌风险生物标志物可供选择，各具有其优势与局限性。广泛应用的障碍包括指南支持不足、临床实践惯性以及有限的临床效用数据。正在进行的多中心试验旨在提供高质量数据，以期纳入指南推荐。
**5. 结论（Conclusions）**
不确定肺结节的分诊数十年来未发生实质性演变。尽管已开发众多风险计算模型，但无一能始终证明优于临床判断。临床医师自然而然地倾向于谨慎处理，导致大量良性结节接受活检，造成不必要的医疗支出与风险。多种生物标志物已被开发用于不确定肺结节的风险分层。尽管大多数显示出良好的诊断准确性，但证明其临床效用的数据有限且质量相对较低。首个旨在评估临床效用的多中心前瞻性随机对照生物标志物试验（ALTITUDE试验）目前正在进行中。随着更多高质量数据的涌现，指南需要明确生物标志物应如何以及何处被纳入不确定肺结节管理算法。目前，生物标志物最适于作为辅助工具；它们并不能取代诊断性影像或其他判定策略的必要性，而仅是作为额外数据点帮助临床医师为患者做出最负责任的诊疗决策。尽管我们身处机器人支气管镜检查和锥形束CT（cone beam CT）成像的时代，但临床医师务必牢记：能够进行某项操作并不意味着应该进行该项操作。
**4. 讨论（Discussion）**
**4.1 生物标志物的必要性（Need for Biomarkers）**
缺乏高敏感性、具有广泛适用性的非侵入性筛查检测可能有助于解释不确定肺结节指南推荐异质性的原因。尽管低剂量CT（Low-Dose CT, LDCT）成像对高危人群（定义为50-80岁当前或既往吸烟者）有益，但其排除了年轻患者及无吸烟史者。此外，LDCT的特异性约为73%，这可能反映了不确定肺结节中频繁的非恶性病因。不幸的是，其中许多结节接受了不必要的侵入性活检，导致与肺癌相关的高额医疗支出。Medicare分析估计，良性结节占肺癌总医疗费用的近40%，包括侵入性操作。这一现状催生了对易于采样、价格低廉且能可靠识别早期肺恶性肿瘤患者的生物标志物的迫切需求，从而降低成本并提供临床获益。
**4.2 关键异同点（Key Similarities and Differences）**
许多已开发的生物标志物整合了生物标本与临床因素，而另一些则仅依赖生物标本。各产品的验证人群略有不同，因此并非每项检测都适用于所有结节。与基于临床因素的风险计算模型（Mayo模型、Brock模型等）类似，必须将正确的生物标志物用于正确的患者。各商业化生物标志物所拥有的数据量及研究的科学严谨性差异显著。重要的是，下文讨论的任何生物标志物均未进行过头对头比较。任何比较性陈述均应被理解为与标准治疗（通常由个体医师层面决定）的比较。
Nodify XL2（Biodesix公司）采用基于血液的两项蛋白（LG3BP和C163A）评估，结合五项临床参数，对8-30 mm肺结节进行再分类。在PANOPTIC（肺结节血浆蛋白组学分类器）研究中，整合模型中LG3BP和C163A蛋白的相对丰度以超过95%的敏感性和阴性预测值区分良恶性肺结节。ORACLE研究以多中心前瞻性非随机方式评估了临床效用：Nodify XL2的使用每检测7名患者可避免1次侵入性操作，实现74%的相对风险降低。这些发现被Kheir等人在多中心回顾性分析中复现。正在进行的ALTITUDE试验将对临床效用进行更严格的评估。
虽已退市，Reveal和Knodule ID（MagArray公司）曾将机器学习技术应用于七种血浆蛋白（癌胚抗原、表皮生长因子受体、中性粒细胞激活蛋白2、前表面活性物质蛋白B、C-X-C基序趋化因子配体10、晚期糖基化终末产物受体及金属蛋白酶组织抑制剂-1）与七个临床变量所组成的阵列，以开发风险再分类模型。该模型将不确定肺结节分为高危和低危类别，改善了真阳性/假阴性比值。Optellum公司采用类似的基于人工智能（Artificial Intelligence, AI）的方法，即肺癌预测卷积神经网络（Lung Cancer Prediction Convolutional Neural Network, LCP-CNN）。该AI计算机工具基于CT扫描结果提供更准确的恶性风险评估，将不确定肺结节的平均诊断区分度从不使用AI工具时的曲线下面积（Area Under the Curve, AUC）0.85提升至0.92（p < 0.001）。在不确定肺结节患者中，计算机辅助诊断还提高了阅片者的平均AUC、敏感 seeded_reader}% ?sit?tθ 3739dues to its nature as an auxiliary diagnostic method.
Among patients with intermediate pulmonary nodules, computer-assisted diagnosis also improved reader average AUC, sensitivity, and specificity for malignancy identification.
An alternative strategy involves identifying tumor-associated antibodies, which may offer higher specificity and the potential for earlier detection through immune signal amplification. Nodify CDT (Biodesix) uses an indirect enzyme-linked immunosorbent assay (ELISA) to evaluate a panel of seven autoantibodies against seven lung cancer-associated antigens. In a recent validation study of 447 patients, Long et al. reported that the autoantibody panel achieved a specificity of 90% (95% confidence interval [CI], 85–93%), with even higher specificity in positron emission tomography (PET)-positive patients. The test demonstrated substantially lower sensitivity (approximately 16%) than Nodify XL2 and a false-positive rate of 10%.
Another method for reclassifying indeterminate pulmonary nodules (IPNs) involves identifying circulating transcriptomic biomarkers, especially microRNAs (miRNAs). These 20–22 nucleotide sequences, located in cancer-associated genes, are increasingly being studied as potential biomarkers for early lung cancer detection, including stage I and II non-small cell lung cancer (NSCLC). Fehlmann et al. demonstrated that a panel of 14–15 miRNAs could differentiate lung cancer with excellent accuracy, with an AUC of 0.965.
Detection of circulating tumor cells (CTCs), circulating tumor DNA (ctDNA), and circulating methylated DNA in blood has mainly been used to identify targetable mutations in thoracic oncology. LungLifeAI (Circulogene), however, is a blood-based assay that employs similar methodology to detect circulating genetically abnormal cells (CGACs). Data show that the test has relatively high specificity (96–100%) but more modest sensitivity (26–80%). The lower sensitivity may be partially mitigated by complementary ctDNA-based tests that detect fragmented tumor DNA in blood.
Newer machine learning models have incorporated more diverse approaches utilizing ctDNA, clinical data, and protein markers to stratify IPNs. ctDNA detection is challenging, however, as only about 30% of early-stage NSCLC cases release ctDNA in detectable amounts. Identifying CTCs can also be challenging because of the mesenchymal transformation of cancer cells. LungLifeAI utilizes immunomagnetic depletion combined with fluorescence in situ hybridization (FISH) to detect chromosomal instabilities in the form of genomic copy number variations (CNVs) in circulating genetically abnormal cells (CGACs) enriched from the peripheral blood. In a cohort of 151 patients, it achieved an AUC of 0.74 (95% CI, 0.66–0.84; p < 0.001), with 67.9% sensitivity and 74.4% specificity, compared with an AUC of 0.52 for the Mayo Clinic Model. Performance was even better in nodules < 2 cm (AUC 0.83), subsolid or nonsolid nodules (AUC 0.90), and stage I disease (AUC 0.80), with comparable performance regardless of histology. Its greatest strength lies in its superior sensitivity in identifying stage I lung cancer, which has the greatest need for additional stratification.
CyPath Lung (BioAffinity) is a sputum-based assay that utilizes automated flow-cytometry to assess for specific antibodies and fluorescent tetra (4-carboxyphenyl) porphyrin (TCPP) that preferentially binds to cancer cells along with machine learning techniques. Lemieux et al. found that while the test fared well with all nodules (80% sensitivity and specificity), it excelled with those < 20 mm in size, where it was 92% sensitive and 87% specific (AUC of 0.94; 95% CI 0.89–0.99).
The Percepta Genomic Sequencing Classifier (GSC) differed from the above tests as the epithelial samples were collected from the main stem bronchus using cytology brushing during bronchoscopy. This test was particularly helpful when bronchoscopy was inconclusive. It used genomic features from whole transcriptome RNA sequencing leveraging 23 genes and clinical factors.a??es [pubReader}% l{tex@ -*
