Abstract 2091: DeepBTS: Prediction of recurrence-free survival of non-small cell lung cancer using time-binned deep neural network

Abstract

Abstract Introduction: Accurate prediction of non-small cell lung cancer (NSCLC) prognosis after surgery is still challenging. Cox proportional hazard model is widely used, but there are some limitations like proportional hazard assumption, computational complexity, and unspecified time component. In this study, we developed the novel neural network models using 30 clinicopathological features of surgically resected NSCLC patients. Materials and Methods: We analyzed a dataset of training cohorts of 1,034 patients and external validation cohort of 303 patients, and compared the novel neural network model and Cox PH model to predict 3-year recurrence of the patient. Results: We developed deep neural network for binned time survival analysis (DeepBTS) algorithm, which use root-mean-squared error (supervised learning model, s- DeepBTS) or negative log-likelihood (semi-unsupervised learning model, su- DeepBTS) as loss function. The su-DeepBTS algorithm showed the improved performance (C-index, 0.7166; AUC, 0.7606) compared to other models (Cox PH: 0.6861 and 0.7185; s-DeepBTS: 0.6926 and 0.7250, respectively). Top 10 features was selected from su-DeepBTS model and su-DeepBTS selector pair: the number of lymph node metastasis, tumor size, histologic subtype, vascular invasion, R0 resection, neoadjuvant /adjuvant treatment, and performance status of patients. These features distinguished low and high-risk group in the training cohort (p = 4.25×10-5) and the validation cohort (p = 1.07×10-11). Conclusions: The su-DeepBTS model using 10 selected features can predict the prognosis of resected NSCLC patients. We plan to diversify the input features by radiologic, pathological imaging and genomic data to integrate multimodal medical information and enhance the performance of this model using deep neural network. Citation Format: Ho Jung An, Bora Lee, Sang Hoon Chun, Ji Hyung Hong, In Sook Woo, Seoree Kim, Joon Won Jeong, Jae Jun Kim, Hyun Woo Lee, Sungsoo Park, Heejoon Jo, Yoon Ho Ko. DeepBTS: Prediction of recurrence-free survival of non-small cell lung cancer using time-binned deep neural network [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2091.