Development of an optimized multi-biomarker panel for the detection of lung cancer based on principal component analysis and artificial neural network modeling

Lung cancer is a public health priority worldwide due to the high mortality rate and the costs involved. Early detection of lung cancer is important for increasing the survival rate, however, frequently its diagnosis is not made opportunely, since detection methods are not sufficiently sensitive and specific. In recent years serum biomarkers have been proposed as a method that might enhance diagnostic capabilities and complement imaging studies. However, when used alone they show low sensitivity and specificity because lung cancer is a heterogeneous disease. Recent reports have shown that simultaneous analysis of biomarkers has the potential to separate lung cancer patients from control subjects. However, it has become clear that a universal biomarker panel does not exist, and optimized panels need to be developed and validated in each population before their application in a clinical setting. In this study, we selected 14 biomarkers from literature, whose diagnostic or prognostic value had been previously demonstrated for lung cancer, and evaluated them in sera from 63 patients with lung cancer and 87 non-cancer controls (58 Chronic Obstructive Pulmonary Disease (COPD) patients and 29 current smokers). Principal component analysis and artificial neural network modeling allowed us to find a reduced biomarker panel composed of Cyfra 21.1, CEA, CA125 and CRP. This panel was able to correctly classify 135 out of 150 subjects, showing a correct classification rate for lung cancer patients of 88.9%, 93.3% and 90% in training, validation and testing phases, respectively. Thus, sensitivity was increased 18.31% (sensitivity 94.5% at specificity 80%) with respect to the best single marker Cyfra 21.1. This optimized panel represents a potential tool for assisting lung cancer diagnosis, therefore it merits further consideration.