Modeling and Quantifying Anticipative Normalization Leakage in Contactless Respiratory Event Detection

This is the authors' abstract. We don't add key points for this paper.

Contactless millimetre-wave radar can register the temporary cessation of breathing without touching the subject. Reported detection results, however, rest on small cohorts, single data partitions and feature pipelines whose normalisation step is computed over the entire recording. That step is invisible in a retrospective analysis and impossible to reproduce in a system that must decide as samples arrive. This study formalises the problem and measures it. A property termed causal admissibility is defined for feature maps, an anticipative normaliser is shown to be a strictly monotone function of the fraction of the recording occupied by the target event, and a leakage magnitude is defined as the paired performance gap between an anticipative pipeline and its causally restricted counterpart. A detector with a causally admissible normalisation stage is then constructed, in which a two-heap streaming quantile replaces the whole-recording statistic at a cost that grows logarithmically with elapsed time, and backward-looking context replaces centred context. Evaluation uses an openly published 60 gigahertz radar dataset of 110 participants containing 220 breath-hold events annotated by a physical push-button. Thirteen classifier families and five feature representations are compared under ten seeds of subject-grouped five-fold cross validation with identical folds, giving fully paired contrasts. The measured leakage magnitude is 0.0065 in area under the receiver operating characteristic curve, present on every one of eighty matched comparisons and large relative to its own variability. Backward-looking context restores discrimination to within the prespecified equivalence margin, leaving the causal-normalisation detector at 0.9956 and statistically equivalent to the retrospective pipeline within 0.005 in area under the curve. The discrimination is high because the task is a controlled voluntary breath hold rather than spontaneous nocturnal apnea, so the accuracy is a substrate for measuring the evaluation artefact and not a clinical claim. Keywords Contactless sensing; respiratory cessation; frequency-modulated continuous-wave radar; anticipative leakage; causal normalisation; streaming quantile; subject-grouped cross validation; reproducible evaluation; subgroup performance audit