Predictive Value of Multi-Parameter Model Incorporating PET-Based Radiomics Features for Survival Prediction of Patients with Diffuse Large B-Cell Lymphoma in Older Patients

Objective To explore the prognostic value of ¹⁸F-FDG PET-based radiomics features by machine learning in diffuse large B-cell lymphoma (DLBCL) in older patients.

Methods A total of 166 DLBCL old patients (≥60 years; 88 males,78 females, age 60-93 years) from March 2011 to November 2019 who underwent pre-therapy ¹⁸F-FDG PET/CT were enrolled in the retrospective study. There were 115 patients in training cohort and 51 patients in validation cohort. The lesions in PET were manually drawn and the obtained radiomics features from patients in training cohort were selected by least absolute shrinkage and selection operator (LASS0), random forest (RF), extreme gradient boosting (Xgboost), and then classified by support vector machine (SVM) to build radiomics signatures (RS) for predicting overall survival (OS). A multi-parameter model was constructed by using Cox proportional hazard model and assessed by concordance index (C-index).

Results A total of 1421 PET radiomics features were extracted and reduced to 10 features to build RS. The univariate Cox regression analysis showed that RS was predictor of OS (hazard ratio (HR)= 5.685, 95% CI: 2.955-10.939; P<0.001).The multivariate model that incorporated RS, metabolic metrics, clinical risk factors, exhibited significant prognostic superiority over the clinical model, PET-based model, and the National Comprehensive Cancer Network International Prognostic Index (NCCN-IPI) in terms of OS: (training cohort: C-index: 0.752 vs 0.737 vs 0.739 vs 0.688; validation cohort: C-index: 0.845 vs 0.798 vs 0.844 vs 0.775).

Conclusions RS could be used as a survival predictor for old DLBCL patients. Furthermor, the multi-parameter model incorporating RS is able to successfully predict patient prognosis.

No relevant conflicts of interest to declare.