Table 7 Relative contribution of each modality to the trimodal LightGBM model, computed using absolute SHAP value aggregation.

From: Multimodal machine learning for 5-year mortality prediction after percutaneous coronary intervention

Modality

Raw sum (%)

Mean-normalized (%)

Visual

27.98

12.32

Textual

48.91

14.35

Tabular

23.11

73.33

  1. Raw sum aggregates contributions across all features within a modality, while mean-normalized adjusts for the number of features to reduce dimension bias.
Back to article page