Abstract Body (Do not enter title and authors here): Introduction: Accurate subtyping of cardiac amyloidosis (CA) into light-chain (AL) or transthyretin (ATTR) forms is essential for targeted therapy but often relies on biopsy or ^99mTc-PYP scintigraphy - tests that are invasive or involve ionizing radiation. Cardiac MRI (CMR) provides rich morphologic and tissue-characterisation data, yet its complex multi-sequence interpretation limits routine subtype assignment.

Research Questions: A deep learning model integrating multimodal CMR (cine, LGE, T1/T2 maps), demographic data, and key laboratory values, including comprehensive serum and urine light chains could enable high-accuracy, non-invasive CA subtype classification with improved interpretability.

Methods: We developed a deep learning model using data from 122 CA patients from the SCMR Registry (61 AL, 61 ATTR; mean age 70.2 ± 11.0 y; 24 % female) confirmed per society guidelines. Sequence-specific encoders included an xLSTM for cine, 3D CNNs for LGE, and 2D CNNs for parametric maps. Demographic data (age as Fourier features, sex as embedding) and labs (light chains, M-protein; continuous as Fourier features, categorical as embeddings) were processed via MLPs. These non-imaging embeddings were integrated with aggregated CMR features using a cross-attention mechanism. Interpretability was provided by (1) Monte-Carlo dropout for uncertainty, (2) modality-gate weights, and (3) Grad-CAM. Five-fold cross-validation (patient-level splits) evaluated performance; operating thresholds were chosen by maximising the Youden index (J = sensitivity + specificity – 1).

Results: Across 5-fold cross-validation, the model achieved a mean ROC-AUC of 0.92 ± 0.05. At the fold-specific Youden thresholds it reached sensitivity 0.82 ± 0.05 and specificity 0.93 ± 0.05 for ATTR. Confidence scores derived from Monte-Carlo dropout averaged 0.68, likely scaled down due to a combination of aggressive dropout during training with cross entropy loss, enabled automatic flagging of uncertain cases while other interpretability methods provided insights into model decision-making.

Conclusions: Integrating CMR, demographics, and light chains into an interpretable multimodal deep-learning framework enables accurate, non-invasive CA subtyping while quantifying predictive confidence. By coupling strong performance with uncertainty-aware triage and transparent saliency cues, the model could reduce reliance on scintigraphy or biopsy and expedite subtype-directed care pending prospective validation.