Machine learning–based sequential analysis for optimal selection between IRD and KRD regimen in multiple myeloma patients.

Abstract

7525 Background: Multiple myeloma (MM) is hematologic malignancy where personalized treatment strategies are essential to improve outcomes. Proteasome inhibitor-based therapy, including ixazomib (IRD) or carfilzomib with lenalidomide and dexamethasone (KRD), is one of the most commonly used regimens for treatment of MM. Early treatment response (ETR) and progression-free survival (PFS) are critically important, as these factors are among the most significant considerations in clinical decision-making for selecting optimal drug regimens in MM treatment. This study aimed to develop a machine learning (ML) model using real-world data (RWD) to predict ETR and PFS in MM patients treated with these two regimens. Methods: This was a retrospective analysis using real-world data (RWD) from 535 MM patients treated with either IRD or KRD regimens. We developed separate ML models for IRD and KRD treatment group to stratify the risk for inadequate treatment response for predicting ETR or PFS outcomes. ETR was assessed based on myeloma protein and the extent of plasma cell proliferation. Patients with complete response (CR), very good partial response (VGPR) were classified as the low-risk (LR) group, indicating optimal treatment response. The other patients were categorized as the high-risk (HR) group, reflecting a suboptimal treatment response. PFS was defined by the standard definition for MM treatment: LR group as at least 2 years of survival without disease progression and otherwise, HR group. ML models were trained using demographic, clinical and genetic data obtained at the initial diagnosis or prescription for IRD or KRD. Additionally, each patient's predicted ETR risk was incorporated as a feature in training the PFS model to enhance its predictive accuracy. Results: The area under the receiver-operating characteristic curve (AUROC) of of the PFS models improved significantly from 0.72 and 0.62 without predicted ETR as feature to 0.81 and 0.85 with ETR incorporated. ML models classified patients as 6 subgroups to suggest optimal drug selection for each group to improve ETR and PFS outcomes. Notably, IRD or KRD given to all patients without considering patient subgroups resulted in a PFS hazard ratio of 5.94 (95% CI: 1.59 – 10.31). Implementation of our ML models might inform drug drug selection in 369 patients (69.0%) among a total of 535 patients. Conclusion: In this study, ML models were developed to predict ETR and PFS in MM patients, facilitating personalized therapy in clinical practice. PFS models demonstrated improved performance by incorporating predicted ETR risk as an additional feature. The ETR-incorporated models with clinical and genetic data stratified patients into low-and high-risk groups proposing optimal treatment strategies to potentially improve PFS in 185 patients (35%). Our findings implicate the potential of RWD and ML to advance precision medicine and improve outcomes through ML-informed treatment decisions.