Figure 6.

SHAP waterfall plot for a correctly classified finger-tapping recording. The model correctly predicted the recording as poor, with extended idle time, long video duration, short tapping duration, and a smaller number of taps (than prescribed) emerging as the strongest contributors. Other features, such as brightness ratio and wrist visibility, had minor effects. The explanation highlights specific, actionable issues that users could correct by rerecording (eg, reducing idle time and completing more taps). SHAP: Shapley additive explanations.