preprint · 2026
one baseline, many trajectories
i clustered baseline imaging measurements into severity classes, built sequential random-forest models with in vitro dormancy features, and used SHAP to examine the features contributing to their predictions.
- each thin line follows one lesion across imaging visits. together, these are 1,193 measured trajectories across 22 regimens. the vertical axis is radiodensity in Hounsfield units (HU), not a model prediction.
- the two colours hold baseline severity fixed: severe and less severe. defining these classes from baseline imaging makes the starting state the reference, rather than grouping by the later response.
- watch the spread over eight weeks, not only the brighter class medians. individual paths overlap and diverge within each class; a common starting category does not mean a common course. choose a regimen below the plot to emphasize its paths and medians while keeping the others faintly in view. all 22 regimens are available as chips, or in a compact menu on smaller screens.
- baseline is shown separately from treatment start; its spacing does not encode elapsed time. subsequent visits are at weeks 2, 4, 6 and 8. lines connect observed visits, and the animation is a reading guide—not continuous measurement. ten missing measurements remain gaps, without dropping their trajectories or filling in values. medians use the available measurements at each visit.
- these are descriptive measurements, not fitted trajectories or a treatment comparison. the sequential models use observed prior imaging, rather than forecasting all eight weeks from baseline alone. model results and their interpretation are in the paper.
data: joshwhiteley/marmoset-paper @ bc4cd45 (radiodensity) · data/