AN AI-POWERED SPATIO-TEMPORAL FRAMEWORK WITH LEAVE-ONE-OUT VALIDATION FOR GENOMIC-INTEGRATED PRECISION DIAGNOSIS AND ADAPTIVE THERAPY IN MULTIPLE MYELOMA

Abstract

Multiple myeloma (MM) is a complex hematologic malignancy diagnosed in over 130,000 individuals globally each year, often presenting challenges in precision care due to fragmented data and rigid risk models. This study introduces an AI-enhanced spatio-temporal framework that integrates genomic, imaging, and longitudinal clinical data to improve diagnosis and treatment responsiveness. Using data from 168 MM patients, including lesion images, biomarkers, and cytogenetic profiles, the model employed UNet segmentation, MST topology extraction, LSTM and Transformer networks, and multimodal attention-based fusion of genomic features such as TP53 deletions and chromosomal translocations. The system achieved 91.6% accuracy in risk stratification, significantly outperforming IMWG standards, and predicted therapy response with 87.2% accuracy and quality-of-life deterioration with an AUC of 0.89. Genomic integration improved predictive performance by 14% over clinical-only models, with strong inter-rater agreement (0.84) and an estimated 18.3% reduction in treatment costs. These findings demonstrate the potential of AI-driven multimodal fusion to deliver clinically interpretable, genome-informed decision support for precision oncology in MM.