Parkinsonian rigidity is conventionally scored by a clinician on a coarse ordinal scale, and models trained to reproduce that scale usually need to be re-calibrated per patient because movement signatures vary so much from person to person. This project learns an ordinal latent representation of rigidity directly from movement data, structured so that the ordering of severity is preserved in the learned space without requiring subject-specific calibration.

The result is a subject-independent estimator: a model trained across a pool of patients that still respects the clinical ordering of rigidity severity when applied to a new, unseen patient.

Tracking individual cells through 3D-plus-time confocal microscopy of developing zebrafish embryos is a core bottleneck in developmental biology research — cell density and imaging noise change constantly as the embryo grows, and most tracking architectures assume a roughly fixed, dense field of objects per frame.

This ongoing project develops a sparse-aware transformer architecture for cell tracking that explicitly accounts for the sparsity and density shifts typical of early embryo development, aiming for more robust lineage tracking across the full developmental sequence.