Beyond high performance, an essential topic of machine learning studies is to dissect the “black box” method and understand how it works. To reveal the regulatory relationships among joints in our model, we further performed SHAP analysis and presented the results as a heatmap of the average absolute SHAP values (Fig. 9a and Table 2). In this heat map, the joints in the column serve as features to predict other joints, whereas the joints in the row are the targets being predicted. The color bars represent their locations in the finger, wrist, or toe. Double lines represent joint space narrowing and a single line represents bone erosion. Higher SHAP values demonstrate more important contributions in predicting the damage scores. As expected, the strongest predictor is generally the joint itself, which is shown as the red diagonal trace in Fig. 9a. Joints from the wrist are often predictable by each other due to their proximity, shown as clustered redness in two submatrices (narrowing and erosion) with orange borders. Intriguingly, we also observed off-diagonal red traces highlighted in four dashed rectangles. For example, the very top dashed rectangle displays the contribution of bone erosion to predict joint narrowing in hand - if a joint has erosion, it is more likely to have narrowing. Altogether, these four rectangles manifest the interconnections between narrowing and erosion in both hands and feet. In addition to the regulation among joints from the same side, we also investigated the contributions of joints from the other side. Again, we observed a stronger diagonal trace, indicating the special and important role of the counterpart joint from the other side in predicting joint damage (Fig. 9b and Table 3). In summary, the SHAP analysis reveals the working mechanisms underlying our machine learning model, especially the multidimensional associations across joint damage types and locations. Consistent with correlation analysis (Fig. 7) and the distinct pattern of joint damage distribution in RA, our method grasps useful information and empowers predictions of joint damage.