Flame-wall interaction (FWI) induces highly localized, reaction-coupled wall heat transfer, which remains challenging to model in large-eddy simulation (LES), particularly at practical grid resolutions. Time-integrated wall heat flux is therefore a relevant metric for thermal loading and design assessments. A key challenge is that DNS-trained models are not guaranteed to remain accurate when transferred to typical LES spatially filtered fields and coarser near-wall grids. To address this gap, the present work targets LES-relevant resolutions by performing a priori tests on spatially filtered DNS fields and by outlining an a posteriori LES validation in a configuration consistent with the DNS. We develop a machine-learning (ML) model for time-integrated wall heat flux during FWI. The inputs are nondimensional features describing flame-wall proximity, local thermochemical state, and flame-wall geometry. Training data are drawn from DNS of premixed methane flames, including head-on and side-wall configurations. The dataset covers pressure-rising constant-volume vessel cases (1D/2D) and constant-pressure 1D cases. Earlier evaluations on 2D/3D turbulent flames showed that the model reproduces DNS trends. In this study, to assess performance at LES-relevant grid resolutions, additional a priori tests are conducted using spatially filtered DNS fields at filter widths representative of typical LES grid spacings. Performance is examined in terms of both trend reproduction and the magnitude of integrated heat flux across filter widths and interaction configurations. In addition, we outline an a posteriori LES validation strategy for the same reacting-flow configuration as the DNS. The LES wall heat flux predicted by ML enables benchmarked against the DNS exact value to evaluate applicability in LES.