Fig. 5: Plasma proteins attributions for cancer and dementia.

The circular SHAP plots of cancer (a) and dementia (b) show the extent to which each protein would affect the proteomic neural network model. Proteins shown in the figure were pre-selected through age- and sex-adjusted CPH models under Bonferroni correction (p-value < 6.84 × 10⁻⁶ = 0.01/1461, p-value was derived corresponding to a two-sided test). The important proteins ranking in the top 1%, as determined by SHAP values, are bolded to emphasize their significance in disease prediction. The width of the range of the bars indicates the extent of proteins’ contribution to disease prediction, with wider bars reflecting a greater contribution. The color of the bars represents the magnitude of plasma proteins, ranging from low (blue) to high (red), as depicted in the color bar above the middle. Deviations toward the center and periphery signify protective and risk contributions, respectively. Noted, four of the top 15 proteins selected by SHAP were not shown in dementia as they did not surpass multiple testing, probably due to their nonlinear effect or high correlations with adjusted covariates. Source data are provided as a Source Data file. CPH model Cox proportional hazard model, SHAP SHapley Additive exPlanations.