Metrics for Distinguishing Biological and Interventional Change in AI Models

Statistical and machine-learning models of longitudinal biological data evaluate change by comparing each new observation against the trajectory implied by prior observations, assuming the process generating that trajectory is stable. We use "data substrate" to mean the underlying structure of the longitudinal data that determines what any such model can recover, independent of its architecture or capacity. When the generating process changes -- whether through a biological transition or through an external intervention -- the prior trajectory ceases to be a valid reference, and extrapolated predictions can be confidently wrong with no internal signal that the reference has failed. A distinct and recognised difficulty is that biological change and interventional change, observed only through serial intertemporal comparison under an assumed trajectory, are readily conflated; existing approaches address this through causal assumptions or hidden-confounder models rather than from the data substrate itself. Here we ask whether the two can be distinguished at the substrate level, and we introduce two subject-level metrics that quantify the geometric signature an interventional change leaves in the data: Curvature Shift, the change in trajectory slope across the event, and Deformation Risk, the departure of post-event observations from the prior-trajectory reference. We evaluate the condition on longitudinal cognitive measurements from 309 human subjects in the Alzheimers Disease Neuroimaging Initiative (ADNI), a large longitudinal dataset containing two distinct, ex-ante-defined regime-change events in the same subjects: a biological transition and an intervention. A model extrapolating the pre-event trajectory assigned the wrong direction of change to roughly two-thirds of post-event observations (post-event sign accuracy 0.341 after the biological event and 0.350 after the intervention, against a chance value of 0.50); only 11% of postbiological-event and 12% of post-intervention readings remained concordant with prior dynamics, and a higher-capacity multilayer perceptron reproduced rather than resolved the error. Curvature Shift was 2.23-fold higher after the biological event (p = 4.4x10-8) and 2.26-fold higher after the intervention (p = 7.4x10-8), and the two metrics were coupled ({rho}= 0.500; 95% CI, 0.407-0.587). Findings replicated on an independent endpoint and survived propensity matching, permutation, 1 and leave-one-out. The metrics detect, per subject, when a fitted models reference has stopped governing the data and whether the departure carries the geometric signature of an interventional change.