Shared　structure　nonlinear　autoregressive　with　exogenous　input　(NARX)　model　is　a　promising　tool　for　exploring　cortical　responses　mechanism　to　external　stimuli,　essential　for　advancing　our　understanding　of　brain　function　and　developing　methods　for　direct　brain　information　encoding.　In　this　paper,　we　proposed　a　two-step　method　to　overcome　limitations　in　existing　method,　which　neglect　data　relationships　and　rely　on　a　greedy　search　for　regression　terms,　leading　to　less　accurate　models.　In　our　approach,　data　from　multiple　trials　are　concatenated,　and　then　the　orthogonal　forward　regression　(OFR)　algorithm　identifies　model　terms　in　first　step,　enhancing　inter-trial　connections　and　establishing　a　preliminary　model　for　each　subject.　Shared　model　terms　across　subjects　are　then　used　to　construct　a　general　target　model.　Next,　non-shared　regression　terms　that　best　represent　population-level　information　are　identified,　using　adaptive　multi-population　genetic　algorithms,　and　use　to　enhance　the　target　models＇　descriptive　power.　Simulations　results　show　significant　competitiveness　in　terms　of　accuracy　as　compared　to　other　state-of-the-art　methods.　When　applied　to　real　electroencephalography　signals　under　mechanical　disturbance,　structural　and　parameter　analysis　revealed　consistent　neural　response　patterns　across　subjects,　with　subject-specific　responses　likely　stemming　from　muscle　feedback.　Frequency　response　analysis　further　suggests　that　the　brain　may　generate　motor　inhibition　signals　based　on　sensory　inputs　to　maintain　a　pre-disturbance　resting　state.　These　findings　provide　valuable　insights　into　cortical　response　mechanisms　and　have　potential　implications　for　future　brain　information　encoding　research.　©　2025　Elsevier　B.V.