Machine-learning-augmented prediction of tumor metabolic burden and survival in multiple myeloma using clinical indicators

Abstract

Abstract Background:PET-CT is a powerful imaging tool that can provide a detailed display of the lesion distribution, quantity, size and metabolic activity in patients with multiple myeloma (MM). It can assist in the early diagnosis of MM, guide tumor staging, evaluate therapeutic effects and prognosis. Despite this, PET-CT examination has some limitations, including high cost, radiation exposure risk, and may be less sensitive than MRI in detecting diffuse lesions in the bone marrow, especially when the lesions are relatively small. Existing studies have shown that there is a correlation between the clinical biological indicators of multiple myeloma and the results of PET-CT. To further clarify the predictive value of clinical indicators for the diagnostic results of 18F-FDG PET/CT, we conducted a retrospective study on 278 patients. This study aims to establish a predictive model by using the PET/CT scan indicators and clinical indicators of MM patients to predict some diagnostic results of 18F-FDG PET/CT. Methods:We retrospectively analyzed 278 patients with multiple myeloma who underwent 18F-FDG PET/CT imaging and clinical laboratory testing. Progression-free survival (PFS) was defined as the time from diagnosis to disease progression or last follow-up. First, we performed univariate and multivariate Cox proportional hazards models to evaluate the prognostic value of SUVmax. SUVmax was found to be significantly associated with PFS. To explore whether SUVmax can be predicted from routine clinical data, we excluded one patient with missing SUVmax and removed the top 20 outliers (SUVmax >18.5), yielding 258 patients (mean SUVmax: 7.29; SD: 3.91). A total of 39 clinical variables were used as predictors. Missing values were not imputed and were handled natively by the XGBoost regression model. SUVmax values were stratified into 10 bins for frequency-based stratified 5-fold cross-validation. Model performance was assessed using RMSE, MAE, MSE, and R². Feature importance was averaged across folds. A final model was trained on 80% of the data and evaluated on a 20% hold-out test set. Findings:SUVmax was independently associated with shorter progression-free survival in Cox regression, highlighting its prognostic value in patients with multiple myeloma. To predict SUVmax without imaging, we trained a machine learning model using routine clinical indicators. The XGBoost model achieved a cross-validated RMSE of 3.731, MAE of 2.864, and R² of 0.031, while the hold-out test set achieved an RMSE of 0.507, MAE of 0.339, and R² of 0.986. The most predictive features included extramedullary disease,soft tissue-related(EM-S),focal lesions, extramedullary disease,bone-related(EM-B), age, total light chain, immunophenotype, CRP, creatinine, α1-globulin, and serum amyloid A. Conclusion These results suggest that SUVmax, a prognostic biomarker for PFS, can be approximated using standard clinical indicators, potentially aiding decision-making in settings where PET/CT is unavailable or impractical.