Decay heat measurements on spent nuclear fuel (SNF) provide a basis for code validation. Their applicability domain (AD) and gaps, which are the focus of this study, are not commonly discussed in the literature. The analyzed validation data are based on measurements at the Clab facility and on calculations using the Polaris and ORIGEN codes of the SCALE code system. Bias-predicting machine learning (ML) models are applied: random forest and weighted k-nearest neighbors. The models weigh the similarity between the cases, expressed using correlations. The learning curves are studied by examining the prediction error versus the sample size and the similarity coefficient. The obtained error reduction at higher similarity coefficients supports the argument that the similarity or correlation is informative. However, a marginal error reduction is expected from increasing the validation data size from its current status. Following this, a validation AD is proposed as a range of SNF characteristics within which the validation data and the ML models are observed and tested. Within the AD, different levels of error, i.e., safety margins and conservatism, were evaluated. Beyond the AD, validation gaps exist. Examination of light-water reactor SNF applications indicates that the validation coverage is absent in both MOX fuel and short cooling, diminishes rapidly at higher burnup for low-enrichment fuel, and extends with burnup for high-enrichment cases. Additional measurements are justified to reduce conservatism or achieve validation coverage in applications. A case study of typical UO₂ and MOX SNF applications is analyzed. It is shown that a few tens of optimally selected measurements from both SNF types are necessary to complete validation coverage in numerous applications.