Dynamic multimodal survival prediction in multiple myeloma integrating gene expression, longitudinal laboratory measurements, and treatment history.

Abstract

Prognostic stratification in multiple myeloma (MM) relies on staging systems fixed at diagnosis, discarding temporal information accumulated during treatment. We developed a dynamic multimodal framework that predicts residual overall survival from observation windows of 1-18 months post-diagnosis. The model integrates DeepInsight-transformed gene expression, longitudinal trajectories of 10 laboratory analytes, and treatment history through missingness-aware gated fusion. On the Multiple Myeloma Research Foundation (MMRF) cohort from the Relating Clinical Outcomes in Multiple Myeloma to Personal Assessment of Genetic Profile (CoMMpass) study (n = 752), five-fold-specific models trained on the development dataset achieved a mean concordance index (C-index) of 0.773 ± 0.024 and 1-year time-dependent area under the receiver operating characteristic curve (AUC) of 0.789 ± 0.021 on a common held-out CoMMpass validation split, outperforming the evaluated survival-learning baselines including DeepSurv, a Cox proportional hazards neural network, and random survival forests. Kaplan-Meier stratification showed significant separation at all primary landmarks (log-rank $P<.001$, hazard ratios 3.46-3.93). A distilled student model retaining only the DeepInsight gene expression representation and five baseline clinical features transferred to an independent microarray cohort (GSE24080, n = 507) without retraining, achieving a C-index of 0.672 and a time-dependent AUC at 1-year of 0.740, supporting cross-cohort transferability in a reduced-input setting. Interpretability analyses recovered ubiquitin-proteasome, endoplasmic reticulum (ER) stress, and Interferon Alpha Response signals consistent with established myeloma biology. These findings support the potential of dynamic multimodal modeling for longitudinal prognostic assessment in MM.