Abstract
390
Aim: To assess 21-month survival prediction modelling and detect promising prognostic markers from in vivo textural parameters derived from [18F]FDG-PET/CT using ensemble machine learning based multi-fold training program and Monte Carlo cross validation in an cohort of patients with treatment naïve pancreas tumors.
Methods: 27 patients (10 female, 17 male; mean age 63.2 years, SD 11.0; range 43-88) with pancreatic cancer who had not received treatment before [18F]FDG-PET/CT were enrolled in this study. Each tumor was characterized by 130 features including in vivo intensity, histogram, shape, textural and joint fusion features as well as general patient demographics (age, gender) and available clinical data (e.g. leukocyte levels, gamma-GT, creatinine). Establishment of predictive models was performed by the utilization of ensemble machine learning approaches in a multi-fold training scheme. Overall 500 folds were established. In each fold 75% of the original data was randomly selected. In each fold 8 machine learning algorithms were executed as presented in [Papp et al]1. The established predictive models were subject to Monte Carlo cross-validation with 1000 iterations. Results: The final predictive models yielded an Area Under the Curve (AUC) of 0.742 and a Sensitivity (SNS) and Specificity (SPC) of 0.853 and 0.631 respectively. Positive Predictive Value (PPV) was 0.698 and Negative Predictive Value was 0.81. Overall feature weights were normalized to 1 before cross-validation. The four most prominent weights were CT-Histogram-Kurtosis (0.219), PET+CT-Fusion-Sum Entropy (0.055), Creatinine (0.048) and PET+CT-Fusion-Correlation (0.043). Conclusion: Ensemble machine learning-based predictive models with Monte Carlo cross-validation can determine 21-month survival in treatment naïve patients with pancreatic tumors with a high SNS, AUC and NPV. 1Papp L, Poetsch N, Grahovac M, et al. Glioma survival prediction with the combined analysis of in vivo 11C-MET-PET, ex vivo and patient features by supervised machine learning. J Nucl Med. November 2017:jnumed.117.202267.