Dynamic biomarker-based machine learning model predicts short-term treatment response in multiple myeloma

Multiple myeloma (MM) is a heterogeneous plasma cell malignancy with variable treatment responses. Conventional prognostic systems such as ISS and R-ISS rely on baseline parameters and fail to capture dynamic changes during therapy. There is an unmet need for models that can predict short-term treatment outcomes (STO) to guide timely clinical decisions. We retrospectively analyzed 662 newly diagnosed MM patients treated between 2017 and 2021. Peripheral blood lymphocyte subsets, cytokine profiles, and bone marrow plasma cell phenotypes (by multiparametric flow cytometry) were assessed at baseline and every two treatment cycles up to Cycle 10. Predictive models were built using Random Under-Sampling Boosting (RUSBoost) and evaluated by cross-validation. Performance was compared with conventional staging systems using F1 score, precision, recall, and accuracy. Cytogenetic abnormalities and ISS/R-ISS classifications did not consistently predict STO beyond Cycle 2. In contrast, dynamic biomarkers—including CD8+ T cells, CD56+ NK cells, lymphocyte counts, and plasma cell surface markers—showed significant associations with treatment responses. The biomarker-based model consistently outperformed conventional staging, reducing false predictions and improving accuracy. At Cycle 4, the model achieved an F1 score of 0.75 versus 0.32 for R-ISS. Both full and feature-selected biomarker sets maintained robust predictive performance across cycles. We developed a dynamic, biomarker-driven machine learning model that accurately predicts short-term treatment response in MM. This approach outperforms conventional staging systems and, importantly, can identify high-risk patients as early as Cycle 2. Such early recognition allows timely therapy intensification or switching before irreversible disease progression, thereby supporting more personalized patient management and potentially improving long-term outcomes.