Predicting The Prognosis Of Multiple Myeloma Using ¹⁸F-FDG PET/CT Based Radiomics Data Implemented With Machine Learning Algorithms

Objective:Multiple myeloma（MM）is the second most common malignant tumor of the hematopoietic system.The prognostic aspects of clinical data have been widely studied,but only a few studies have explored the role of radiomics features based on dual-modalities and multiple algorithms of¹⁸F-FDG PET/CT in the prognosis of MM patients.This study aims to delve deeper into the role of radiomics data in the prognosis of MM.Methods:The 121 patients with MM who underwent ¹⁸F-FDG PET/CT exams at our facility between February 2014 and October 2022 confirmed MM by pathological biopsy were initially included in this retrospective study.However,98 patients were ultimately enrolled in this study after meeting the inclusion and exclusion criteria.Clinical data were obtained from the hospital’s examination and medical record systems,while imaging data were exported from the Med Ex system in DICOM format and then semi-automatically delineated using the free and open-source software LIFEx to evaluate and define regions of interest.Finally,quantitative radiomics features were extracted in batches using the Py Radiomics software.This retrospective study extensively explores the discrimination abilities and clinical decision-making abilities for prognostic of different combinations of radiomics data,clinical data,as well as six survival machine learning algorithms,including Cox proportional hazards model（Cox）,linear gradient boosting models based on Cox’s partial likelihood（GB-Cox）,Cox model by likelihood based boosting（Cox Boost）,generalized boosted regression modelling（GBM）,random forests for survival model（RFS）and support vector regression for censored data model（SVCR）.And the model evaluation methods include Harrell concordance index,time dependent receiver operating characteristic（ROC）curve,and decision curve analysis（DCA）.Results:In this study,we ultimately confirmed five PET-based radiomics features,four CT-based radiomics features,and six clinical data features,including high-risk cytogenetic status,RISS staging,red blood cell count,albumin level,B2M,absolute neutrophil count,which were significantly associated with PFS,and were incorporated into the model construction.Among them,in the validation group,the best model was the RSF algorithm,a combination of clinical data and PET-based radiomics features（average C-index:0.880,95%CI:0.878-0.881）,while in the training group,the GBM algorithm showed the highest accuracy in a combination model of clinical data and CT-based radiomics features（average C-index:0.961,95%CI:0.960-0.962）.Except for the PET-based radiomics model constructed by the SVRC algorithm,which showed no statistical difference compared to the model based solely on clinical data,the combinations of various modes of other algorithms with clinical data showed significant higher predictive ability compared to clinical data alone（P<0.001）.Time-dependent ROC analysis showed that adding PET or CT features to clinical data during follow-up could greatly improve the performance of predicting prognosis for each algorithm.The performance of the PET-based single-modality prognosis model was superior to the baseline model based on clinical parameters and the CT-based single-modality radiomics model.Conclusion:The radiomics model based on¹⁸F-FDG PET/CT images and machine learning algorithms can significantly improve clinical progression prediction and increase clinical decision-making efficiency for multiple myeloma,providing a prospect for accurate prognosis stratification in clinical treatment.At the same time,it also provides a new direction for the application of radiomics in multiple myeloma.