Reliable　intensity　measures　(IMs)　are　essential　for　machine　learning　(ML)-based　structural　seismic　damage　assessment　because　they　capture　critical　ground　motion　characteristics　relevant　to　structural　damage.　The　inherent　randomness　of　ground　motions　and　the　structural　diversity　of　urban　buildings　require　a　subset　of　IMs　instead　of　a　single　IM　for　large-scale　seismic　damage　assessments　to　enhance　the　predictive　accuracy　of　ML　models.　A　comparative　study　of　filter　and　wrapper　methods　was　conducted　to　identify　optimal　subsets　of　IMs　tailored　for　reinforced　concrete　(RC)　frame　structures　in　urban　building　clusters.　The　process　involved:　1)　constructing　a　structural　damage　database　for　typical　urban　RC　frames　through　seismic　response　analyses　on　established　simplified　numerical　models;　2)　establishing　a　comprehensive　set　of　IMs,　comprising　34　IMs　proposed　by　the　authors　and　49　collected　from　literature;　3)　evaluating　individual　IMs　using　filter　methods　based　on　criteria　such　as　correlation,　efficiency,　practicality,　and　proficiency,　selecting　top-ranked　IMs,　and　organizing　them　into　distinct　subsets;　4)　using　wrapper　methods,　such　as　forward　selection,　backward　elimination,　and　genetic　algorithm,　to　initially　select　subsets　of　IMs,　which　were　iteratively　refined　based　on　the　performance　of　ML　models　to　achieve　optimal　accuracy;　5)　comparing　the　performance　of　the　subsets　derived　from　filter　and　wrapper　methods,　identifying　those　with　acceptable　accuracy　and　the　smallest　number　of　IMs;　6)　validating　the　bestperforming　subsets　through　seismic　damage　assessments　of　three　campus　buildings.　The　framework　was　demonstrated　using　extreme　gradient　boosting　(XGBoost)　and　random　forest　as　representative　ML　algorithms.　The　performance　of　filter　and　wrapper　methods　was　evaluated　across　various　types　and　scales　of　IM　sets.　Applicable　scenarios　for　each　method　were　analyzed　considering　predictive　accuracy　and　efficiency,　and　critical　subsets　of　IMs　were　recommended　for　ML-based　seismic　damage　prediction　of　urban　RC　frame　structures.　The　results　provide　reliable　and　efficient　decision　support　for　post-earthquake　damage　assessment　of　urban　buildings,　as　well　as　valuable　references　for　fragility　analysis　of　individual　structures.