We audited 1,057 repetitions from ten REHAB24-6 subjects, separating evaluation aggregation, subset structure, and temporal representation. The manual‑subset kNN gain drops from 0.055 for pooled out‑of‑fold AUROC to 0.020 for equal‑weight within‑person AUROC; both confidence intervals include zero, indicating no statistically reliable improvement.
Among 1,000 dimension‑matched random maps, only 14 match or exceed the manual pooled result. When bilateral structure and trunk inclusion are also matched, the count rises to 145. RBF‑SVM retains a positive within‑person gain, whereas logistic regression and a random‑convolution baseline show negative point gains under the same estimand. Sequence‑order and paired‑seed controls further qualify the interpretation.
This exploratory audit demonstrates that joint‑selection claims require explicit estimands and structurally appropriate controls; it does not introduce a new algorithm nor establish clinical benefit.
Review