18F-FDG PET radiomics score construction by automatic machine learning for treatment response prediction in elderly patients with diffuse large B-cell lymphoma: a multicenter study.

Abstract

Purpose: To explore the development and validation of automated machine learning (AutoML) models for ¹⁸F-FDG PET imaging-based radiomics signatures to predict treatment response in elderly patients with diffuse large B-cell lymphoma (DLBCL).

Methods: A retrospective analysis was conducted on 175 elderly (≥ 60 years) DLBCL patients diagnosed between March 2015 and March 2023 at two medical centers, with a total of 1010 lesions. The baseline PET imaging-based radiomics features of the training cohort were processed using AutoML model AutoGluon to generate a radiomics score (radscore) and predict treatment response at the lesion and patient levels. Furthermore, a multivariable logistic analysis was used to design and evaluate a multivariable model in the training and validation cohorts.

Results: ROC curve analysis showed that the radscore generated by AutoML exhibited higher accuracy in predicting treatment response at the lesion level compared to metabolic parameters (SUVmax, MTV, and TLG) in both the training group (AUC: 0.791, 0.542, 0.667, 0.651, respectively) and the validation group (AUC: 0.712, 0.616, 0.639, 0.657, respectively). Multivariable logistic analysis indicated that NCCN-IPI (OR = 5.427, 95% CI: 1.163-25.317), BCL-2 (OR = 3.714, 95% CI: 1.406-9.816), TMTV (OR = 4.324, 95% CI: 1.095-17.067), and avg-radscore (OR = 3.176, 95% CI: 1.313-7. 686) were independent predictors of treatment response. The multivariable model comprising NCCN-IPI, BCL-2, TMTV, and avg-radscore outperformed conventional models and clinical-pathological models in predicting treatment response. (P<0.05).

Conclusion: The radscore generated by AutoML can predict the treatment response of elderly DLBCL patients, potentially aiding in clinical decision-making.