1550 Background: Low-dose computed tomography (LDCT) is an effective approach for lung cancer screening of high-risk patients with pulmonary nodules, however with varying false positive rates depending on the somewhat subjective judgement of the practice professional. Artificial intelligence derived from machine learning of comprehensive patient profiles, including multi-omics and clinical data, has the potential to provide more objective assessment of patient’s risk in order to aid clinician’s decision making. We have developed a multi-analyte algorithm-based assay (MAAA) that incorporates ctDNA mutation, ctDNA methylation, and protein biomarker profiles evaluated through non-invasive blood-based testing, as well as patient’s clinical information, to improve the diagnostic efficacy of lung cancer. Methods: 98 high-risk patients with pulmonary nodules were enrolled in two independent cohorts (68 for training/testing and 30 for independent validation). The malignancy of the pulmonary nodules were established through pathology of surgical-removed nodules. Prior to surgery, each patient was also subject to cell-free DNA-based sequencing for DNA mutation and DNA methylation profiling, as well as serum protein biomarker profiling. On the training/testing patient cohort, machine-learning-based predictive models were first built for malignancy status prediction based on each type of molecular or clinical features. A final ensemble model was then constructed to incorporate the measurements based on molecular and clinical markers to provide the ultimate recommendation on the malignancy of the pulmonary nodule. The performance of each individual model and the final ensemble model was benchmarked on the training/testing cohort, and also validated on the independent validation cohort. Results: On the 30-patient independent validation cohort, individual prediction models based on clinical information, protein marker, ctDNA mutation, and ctDNA methylation profiles achieved predictive AUC of 0.59, 0.48, 0.71, and 0.84, respectively. The final ensemble model achieved predictive AUC of 0.86, which has strongly indicated that an integrative, algorithm-based approach of multi-analytic molecular and clinical profiles greatly outperforms any single-analytic profiling. Conclusions: Multi-analyte algorithm-based approach can be utilized to assist in lung cancer screening for patients with pulmonary nodules. It has demostrated a high accuracy through independent validation, and has outperformed any single-analyte testing in our study.
New Subsolid Pulmonary Nodules in Lung Cancer Screening: The NELSON Trial
