Geometric Characterisation and Structured Trajectory Surrogates for Clinical Dataset Condensation
The study addresses limitations in trajectory-based clinical dataset condensation by introducing Bézier Trajectory Matching (BTM). BTM replaces standard stochastic gradient descent paths with structured quadratic Bézier curves. This results in a more organized, lower-rank supervision signal that better fits the constraints of fixed synthetic datasets, leading to improved predictive performance, especially in rare clinical scenarios.