Interpreting Public-Dataset SOH Benchmarks: A Systematic Review of Evaluation Protocols and Reproducibility Evidence

Studies that use the same public lithium-ion battery dataset do not necessarily evaluate the same state-of-health (SOH) estimation task; reported performance depends on how the experiment is defined. We mapped 408 studies and then examined a nested set of 168 with SOH-REF. This subset was not prespecified as representative, so its percentages are descriptive only. All four Core interpretability elements could be identified in 95/168 studies (56.5%). Same-cell and unseen-cell tests addressed different generalization settings. Preprocessing fitting scope was assessable in 40/168 studies (23.8%), whereas leakage risk was assessable in 64/168 (38.1%). Uncertainty evidence was uncommon, and explicit article-reported public implementation evidence was rare. These classifications reflect what the papers reported; NOT_ASSESSABLE does not indicate poor methodology or establish whether leakage was present. An accuracy value is therefore most informative when accompanied by the health-target definition, physical train-test relationship, preprocessing fitting scope, metric context, and aggregation procedure. More explicit reporting of these elements would make public-dataset SOH benchmarks easier to interpret and reuse without imposing a single evaluation design.