Identity-Corrected Embedding Deltas for Lung-Nodule Progression

Frozen 3D CT foundation models such as FMCIB are increasingly used as off-the-shelf feature extractors, and their natural longitudinal biomarker, following delta-radiomics, is the element-wise difference between a follow-up and a baseline embedding. What this "embedding delta" actually captures, however, remains poorly understood. We study it for predicting index-nodule progression in the National Lung Screening Trial and report two findings. First, the delta is a genuine progression biomarker when the lesion is well localized: a single-timepoint embedding is near chance (AUC ∼0.56), whereas the delta reaches 0.73, and it is not a size proxy, since no diameter feature enters the predictor. Second, using a joint–individual decomposition (JIVE), we identify a compact patient-identity factor intrinsic to these embeddings. This factor re-identifies the same patient across a one-year interval at 690× chance and, because inter-scan drift leaves it only partially cancelled, still occupies 37% of the delta's variance. Projecting this identity subspace out of the delta, an unsupervised and label-free step, raises progression AUC from 0.73 to 0.81. The patient-identity factor is thus both a re-identification caution for released "de-identified" embeddings and a practical de-noising tool for longitudinal biomarkers.