Abstract 1137P
Background
Advances in precision oncology results in an increased request for predictive tools able to select and stratify patients for treatments. In this contest machine-learning radiomics is a promising approach to improve the clinical management of non-small cell lung cancer (NSCLC). We aimed to use machine learning to evaluate the prognostic role of radiomic feature-based (18F) fluorodeoxyglucose (FDG) positron emission tomography (PET) imaging in NSCLC patients.
Methods
321 NSCLC patients that underwent PET/CT with FDG before any therapy. Patients were divided in two groups: early stage group (n=79) and advanced stage (n=242). A total of 42 radiomics features and standardized uptake values (SUVmax) were extracted from PET studies. Feature selection occurred with a 3-step process consisting of: 1) Spearman’s rank correlation versus clinical stage; 2) least absolute shrinkage and selection operator (LASSO) regression model; 3) evaluation of model predictive performance with bootstrap resampling and area under the curve (AUC).
Results
The variable combination that best distinguished the two groups of patients included 8 textural features and SUVmax. All considered statistical features (specificity, sensitivity, Chi-Square, relative risk and 95% confidence interval, 95% CI) indicate a higher statistical significance of the combined model vs SUVmax. This combined 9-parameter model predicted progression-free survival (P=0.0006 and P=0.01) and overall survival better than SUVmax alone (P=0.0003 vs P=0.08).
Conclusions
These data indicate that machine learning radiomic analysis can improve subjects stratification in NSCLC patients and can be used to predict survival in NSCLC with greater statistical power than SUVmax alone.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Andrea Ciarmiello.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.