XLumi: Causal and Continual Learning with Hierarchical Explanations for Hematology Diagnosis

Hematological diagnosis requires integrating diverse data modalities—bloodsmears, flow cytometry, genomics, and clinical narratives—while adapting to evolv-ing distributions and providing interpretable decisions. We introduce XLumi, amulti-modal diagnostic framework that addresses three key challenges: (1) learningcausal relationships across heterogeneous modalities rather than spurious correla-tions, (2) continually adapting to new disease patterns without catastrophic forget-ting, and (3) generating hierarchical explanations aligned with clinical reasoning.Our approach combines a causal multi-modal fusion transformer with biologically-constrained structure learning, an episodic memory network for continual adapta-tion, and a hierarchical explanation generator. Evaluation on benchmark datasetsdemonstrates strong diagnostic performance while maintaining interpretability andadaptation capabilities suitable for clinical deployment.