Michael Baad MD, Presenter: Nothing to Disclose

Yisheng Chen, Abstract Co-Author: Nothing to Disclose

Yonglin Pu MD, PhD, Abstract Co-Author: Researcher, Eli Lilly and Company Researcher, General Electric Company

Risk stratification in NSCLC is currently accomplished through the TNM clinical staging system, which fails to include many variables that have been shown to have prognostic significance independent of stage, such as FDG-PET tumor measurements. Here we use artificial neural networks (ANNs) to compare the prognostic performance after the inclusion of these imaging and clinical variables to that of the TNM clinical stage alone.

Using comprehensive clinical and imaging data of 328 consecutive NSCLC patients with a baseline PET/CT, three ANN models were constructed with incremental increases in input variables as follows: 1) a baseline model consisting of only the clinical TNM stage and “censored variable”; 2) a reduced model consisting of all variables in the baseline model with the addition of clinical variables including gender, histology type, surgery and chemotherapy type; and 3) a full model consisting of all the variables in the reduced model with the addition of FDG-PET measurements including ln(SUVmaxWB), and MTVWB. The coefficient of determination (R2) and root mean square error (RMSE) were then calculated between predicted and observed survival, with 10% of cases held for cross validation. The same variables were then analyzed by multiple linear regression (MLR).

The R2/RMSE between predicted and observed overall survival improved with the addition of clinical variables, from 0.605/18.5 using only the TNM clinical stage to 0.748/13.8 with the addition of clinical variables. Addition of FDG-PET measurements resulted in even greater performance, with an R2/RMSE of 0.781/13.8. The same trend was found using MLR analysis with R2/RMSE of 0.660/17.4, 0.642/17.8 and 0.581/19.1 in the full, reduced and baseline models respectively.

ANN models can be used to overcome the limitations of the current TNM staging system for better predicting patient’s prognosis in NSCLC by combining the value of multiple prognostic variables. The inclusion of clinical and imaging variables, such FDG-PET measurements, into the models resulted in incremental improvements in performance over the TNM clinical stage alone.

ANN models can improve survival prediction in NSCLC by including prognostic factors currently not included in the TNM staging system, such as FDG-PET measurements. 

Baad, M, Chen, Y, Pu, Y, Application of Artificial Neural Networks for Predicting Survival Time of Patients with Non-small Cell Lung Cancer Using FDG-PET and Other Clinically Available Prognostic Factors.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14045563.html