Fig. 9: Pipeline for constructing UKB-MDRMF.

First, a time alignment process is applied to ensure that disease occurrences post-date the baseline data. The red-shaded section at the bottom illustrates the alignment process for different individuals, where the red dashed line represents the enrollment time. Features such as basic, lifestyle, and other characteristics are collected at the time of enrollment. Phecodes recorded before enrollment are marked in gray and treated as missing values during model training, ensuring they are not used for training purposes. After integrating Phecode data from multiple sources, only the earliest occurrence of the same Phecode post-enrollment is retained. Next, various multi-disease prediction and risk assessment models are applied for comprehensive evaluation. These models are trained separately using distinct loss functions. Finally, model interpretability analysis is performed, incorporating associations between different diseases and risk factors for integrated analysis. Variable importance results are derived from all model weights, while multimorbidity relationships are inferred from the embeddings in the penultimate network layer. Icons provided by Icons8 (https://icons8.com).