A framework for integrating MR imaging and clinical tests for prediction of outcomes in multiple myeloma treatment

Abstract Multiple myeloma (MM) is the second most common hematologic malignancy (after non-Hodgkin lymphoma) in the U.S. Over the past decade, despite the advent of novel therapies that has improved the life expectancy of patients with MM, MM is still incurable. Patient outcome varies widely with overall survival (OS) ranging from

less than 1-year to greater than 10-years. The large variability of patient overall survival resulting from the high heterogeneous microenvironment of MM heightens the needs of risk-adapted treatment for long-term control, which are largely relied on accurate assessment of tumor evolution and survival prediction. The International

Myeloma Working Group (IMWG) guideline and several other different prognostic models have been used to assess treatment outcomes. These systems utilize different subsets of laboratory values from clinic tests as assessment criteria, which may not indicate the full extent of tumor burden. MRI is the gold standard imaging

modality for detection of bone marrow involvement. It provides global representation of the tumor burden, even for patient with non-secretory or hypo-secretory MM. We hypothesize that integrating multimodal information delivered by time serial noninvasive MR imaging and clinical tests will provide more robust and accurate

modeling of MM evolution during and after treatment, and thus can predict treatment outcomes more reliably at an early time. Our goal is to develop a framework of decision support system (DSS) that integrates MR images and clinical tests to assist physicians in making informed decisions on tailoring treatment for an individual MM

patient. The proposed study aims to (1) Investigate the effectiveness of MRI radiomic features in characterizing MM infiltration and their correlations with treatment outcomes, (2) Investigate the circumstances how the combination of MRI radiomics and clinical tests complements each other and is more powerful in early prediction

of treatment outcomes. To achieve these aims, we will retrospectively collect a large data set of MM cases with baseline and follow-up MRI scans and clinical tests for development and validation of the DSS system. We will utilize advanced artificial intelligence (AI) techniques including recurrent neural networks and reinforcement

learning to train predictive models with time serial MRI and clinical data acquired at multiple time points, which allows the trained models to be deployed flexibly to each individual time point for treatment assessment to facilitate early prediction. Furthermore, these advanced AI techniques can learn intertemporal dependencies

from available data to compensate for the missing data or exams at single time points in serial exams which may occur in real-world clinical situation. With a flexible modular architecture design of our DSS system, we will investigate the inter- and intra-relationships between MRI and laboratory variables for better understanding of

their complementarity in characterizing tumor burden during MM treatment and their association with treatment outcomes. As a final step, using an independent multi-institutional test set, we will validate the performance of the developed DSS in predicting the recurrence risk at an early time point during treatment.