Fig. 3: Population-level variable and time window contributions to expected 6-month functional outcome output at high-magnitude transitions.
From: Mining the contribution of intensive care clinical course to outcome after traumatic brain injury
TimeSHAP values are interpreted as contributions of variables or time windows towards the difference in a patient’s expected 6-month functional outcome output from that of the average patient (Supplementary Fig. 13). a TimeSHAP values of the 20 highest-contribution static (left) and 20 highest-contribution dynamic (right) variables. The variables were selected by first identifying the ten variables with non-missing value tokens with the most negative median TimeSHAP values across the population (above the ellipses) and then, amongst the remaining variables, selecting the ten with non-missing value tokens with the most positive median TimeSHAP values (below the ellipses). Each point represents the mean TimeSHAP value for a token across an individual patient’s high-magnitude transitions. The colour of the point represents the relative ordered value of a token within a variable, and for unordered variables (e.g., employment status before injury), tokens were sorted alphanumerically (the sort index per possible unordered variable token is provided in the Supplementary Note 1). Green points represent variable tokens that are not missing but explicitly encode an unknown value (i.e., GCS motor score untestable due to sedation). New variable abbreviations include deep vein thrombosis (DVT), the fraction of inspired oxygen (FiO2), partial pressure of oxygen (PaO2), and unfavourable outcome (UO) as defined by GOSE ≤4 at 6 months post-injury. b TimeSHAP amplitude distributions of 2-h time windows leading up to high-magnitude transitions. The width of violin plots is scaled for each time window, but the width of the points inside them demonstrates relative frequency across the windows.
