Raimundo Bernardo S, Leitão Pedro M, Vinhas Manuel, Pires Maria V, Quintas Laura B, Carvalheiro Catarina, Barata Rita, Ip Joana, Coelho Ricardo, Granadeiro Sofia, Simões Tânia S, Gonçalves João, Baião Renato, Rocha Carla, Alves Sandra, Fidalgo Paulo, Araújo Alípio, Matos Cláudia, Simões Susana, Alves Paula, Garrido Patrícia, Pantarotto Marcos, Carreiro Luís, Matos Rogério, Bárbara Cristina, Cruz Jorge, Gil Nuno, Luis-Ferreira Fernando, Vaz Pedro D
Unidade de Pulmão, Centro Clínico Champalimaud, Fundação Champalimaud, 1400-038 Lisboa, Portugal.
Departamento de Engenharia Electrotécnica e de Computadores, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, Quinta da Torre, 2829-516 Caparica, Portugal.
Cancers (Basel). 2025 May 16;17(10):1685. doi: 10.3390/cancers17101685.
Lung cancer (LC) is the leading cause of cancer-related deaths worldwide. Effective screening strategies for early diagnosis that could improve disease prognosis are lacking. Non-invasive breath analysis of volatile organic compounds (VOC) is a potential method for earlier LC detection. This study explores the association of VOC profiles with artificial intelligence (AI) to achieve a sensitive, specific, and fast method for LC detection. Exhaled breath air samples were collected from 123 healthy individuals and 73 LC patients at two clinical sites. The enrolled patients had LC diagnosed with different stages. Breath samples were collected before undergoing any treatment, including surgery, and analyzed using gas chromatography coupled to ion-mobility spectrometry (GC-IMS). AI methods classified the overall chromatographic profiles. GC-IMS is highly sensitive, yielding detailed chromatographic profiles. AI methods ranked the sets of exhaled breath profiles across both groups through training and validation steps, while qualitative information was deliberately not taking part nor influencing the results. The K-nearest neighbor (KNN) algorithm classified the groups with an accuracy of 90% (sensitivity = 87%, specificity = 92%). Narrowing the LC group to those only in early-stage IA, the accuracy was 90% (sensitivity = 90%, specificity = 93%). Evaluation of the global exhaled breath profiles using AI algorithms enabled LC detection and demonstrated that qualitative information may not be required, thus easing the frustration that many studies have experienced so far. The results show that this approach coupled with screening protocols may improve earlier detection of LC and hence its prognosis.
肺癌（LC）是全球癌症相关死亡的主要原因。目前缺乏能够改善疾病预后的有效早期诊断筛查策略。对挥发性有机化合物（VOC）进行非侵入性呼气分析是一种早期检测肺癌的潜在方法。本研究探索了VOC谱与人工智能（AI）之间的关联，以实现一种灵敏、特异且快速的肺癌检测方法。在两个临床地点收集了123名健康个体和73名肺癌患者的呼出气样本。纳入的患者被诊断为不同阶段的肺癌。在接受任何治疗（包括手术）之前采集呼气样本，并使用气相色谱-离子迁移谱联用仪（GC-IMS）进行分析。人工智能方法对整体色谱图进行分类。GC-IMS高度灵敏，可产生详细的色谱图。人工智能方法通过训练和验证步骤对两组的呼出气谱集进行排序，而定性信息未参与也未影响结果。K近邻（KNN）算法对两组的分类准确率为90%（灵敏度 = 87%，特异性 = 92%）。将肺癌组缩小到仅为早期IA期患者时，准确率为90%（灵敏度 = 90%，特异性 = 93%）。使用人工智能算法评估整体呼出气谱能够检测肺癌，并表明可能不需要定性信息，从而缓解了许多研究至今所经历的困扰。结果表明，这种方法与筛查方案相结合可能会改善肺癌的早期检测，进而改善其预后。
