Unmasking Dementia Detection by Masking Input Gradients: A JSM Approach to Model Interpretability and Precision