Authors
Natalie Lui, Nien Wei, Winston Trope, Shannon Nesbit, Prasha Bhandari, Chin-Hui Lee, Hu Hu, H. Henry Guo, Douglas Z. Liou, Joseph B. Shrager, Leah Monique Backhus, Mark F. Berry, Eric Yang
Stanford University, Stanford, CA; Auspex Diagnostics, Warren, NJ; Georgia Institute of Technology, Atlanta, GA
Abstract
Background: Five-year survival for stage I-II lung cancer is quite low even after complete surgical resection. Current guidelines recommend adjuvant treatment only for selected patients with stage II or higher disease. A prediction model that identifies patients at high risk of recurrence who may benefit from adjuvant treatment is greatly needed. Many existing prediction models include a small number of genes that were found to be significant in previous studies. We propose using artificial intelligence to analyze a microarray of > 20,000 well-annotated genes to create a model that predicts recurrence after surgical resection of stage I-II lung cancer.
Methods: We identified 275 patients who underwent surgical resection for pathologic stage I-II lung adenocarcinoma or squamous cell carcinoma from 2009 to 2019 in our institution’s prospective surgical database. We excluded patients who had follow up time less than 3 years or received adjuvant therapy and had not had a recurrence, as well as patients with missing specimen blocks. Patient characteristics and recurrence information were obtained from chart review. The patients were divided into training (192 patients) and validation (83 patients) cohorts, and the recurrence status for the validation cohort was initially blinded. Gene expression levels were generated using Clariom S human array (ThermoFisher) from 10um sections cut from the formalin-fixed, paraffin-embedded surgical specimen blocks. The artificial intelligence algorithm Support Vector Machine (SVM) was used to create a prediction model for recurrence using the gene expression and recurrence status of the patients in the training cohort. The model was then tested on the validation cohort using Kaplan-Meier analysis and the area under the receiver operator curve (AUROC).
Results: The recurrence prediction model separated the validation cohort into 15 (18.1%) patients in the high-risk group and 68 (81.9%) patients in the low-risk group. Kaplan-Meier analysis showed the five-year disease-free survival was significantly higher in the low-risk group compared to the high-risk group (86 vs. 50%, HR = 4.41, p = 0.0025). The AUROC for predicting recurrence was 0.744.
Conclusions: Our model uses artificial intelligence to successfully predict recurrence after surgical resection for stage I-II non-small cell lung cancer. With an AUROC of 0.744, our model outperforms previously described models with AUROC up to 0.6. Our model separates patients into high-risk and low-risk groups, which will make management decisions clearer compared to other models that also include an intermediate-risk group. Patients in the low-risk group had 86% five-year disease-free survival; patients in the high-risk group had 50% five-year disease-free survival and may benefit from increased postoperative surveillance or adjuvant therapy.