This study presents an adaptive metamodel approach, assisted by a Deep Neural Network (DNN), for damage detection in steel frames based on incomplete frequency and mode shape data with limited training datasets. The proposed method integrates model order reduction (MOR) and a multi-stage process to enhance efficiency and accuracy. Initially, the Modal Strain Energy Change Ratio (MSECR), calculated from incomplete modal data, is employed to eliminate low-risk damage candidates by leveraging a second-order Neumann series expansion-based MOR (NSEMR-II) technique. This significantly reduces the neural network architecture of the DNN model used in subsequent stages. The DNN is trained on frequencies and mode shapes simulated using the Finite Element Method (FEM), corresponding to measured degrees of freedom (DOFs). Iteratively refining damage candidates through a damage threshold, the method improves diagnostic accuracy while maintaining low computational demands and requiring only moderately sized datasets. The simplified DNN models effectively identify both the location and severity of damage using data from limited sensors, even under high noise conditions. Numerical examples on steel frame structures validate the approach’s efficiency and practicality for structural health monitoring applications.