Fig. 4: Evolution of features.
From: Diagnostic framework to validate clinical machine learning models locally on temporally stamped data
The heatmap shows the evolution of the top 15 features (ranked by Shapley values), standardized, and organized into monthly segments. Feature importance is determined using the training set and evolution is monitored beyond this period (white vertical line). For illustration, the white rectangle depicts the row-normalized average albumin serum result of patients initiating systemic therapy in April 2015. Changes in color saturation depict variations in the standardized means (continuous variables) and frequencies (binary variables) of a feature and thus highlight shifts in practices, usage of procedure/diagnosis/laboratory units or distributions of the patient population. Features that transition to dark purple hue indicate their diminishing usage and potential discontinuation (categorical variables) or that the mean has decreased (continuous variables).
