<p>Prompt gamma neutron activation analysis (PGNAA) is widely employed for the rapid, non-destructive, and multi-element analysis of large-volume cementitious waste forms. However, the quantitative determination of low-concentration toxic elements, such as Cd and Hg, within these matrices is often constrained by limitations in conventional methods, particularly inadequate correction for neutron self-shielding and matrix effects. To enhance the predictive accuracy for Cd and Hg, this study proposes an integrated quantitative analysis approach combining PGNAA with deep learning. Specifically, this study develops a deep learning model based on a Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory, incorporating a Feature Extraction Synergy (FES) module, with the full 4096-channel gamma-ray spectrum obtained from Geant4 Monte Carlo simulations as input. The FES module is composed of two serial submodules: the Fourier Spectral Feature Extraction (FSFE) submodule, which combines Fourier transform with deep convolution to capture global frequency domain features, and the Multiscale Higher-Order Feature Fusion (MHOFF) submodule, which employs layered linear transformations and interactive mechanisms to enhance local detail representation. Together, FSFE and MHOFF enable collaborative modeling of global spectral trends and localized features. Furthermore, model hyper-parameters are fine-tuned using an Improved Harris Hawks Optimization (IHHO) algorithm. Testing results demonstrate that the proposed model achieves exceptional predictive performance: coefficients of determination (R²) of 0.9998 for Cd and 0.9719 for Hg. Compared to the best-performing traditional CNN model, the root mean square error (RMSE) is significantly reduced by 36.38% and 48.38% for Cd and Hg, respectively. This approach provides a promising methodological framework for the quantitative analysis of trace-level toxic elements in cement-based solidified waste materials.</p>