Mental-health forecasting from social-media data is a complex multimodal challenge involving temporal, textual, and relational information. This study presents a hybrid two-stage framework that integrates a Long Short-Term Memory (LSTM)-based graph ensemble with an Ensemble Deep Kernel Learning (EDKL) meta-model to predict depressive risk and emotional trajectories within online social networks. In Stage 1, user-level representations are encoded using an LSTM encoder combined with multiple graph neural backbones, including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Graph Transformer Networks (GTN). Their outputs are stacked and calibrated via logistic regression to produce reliable depression-risk probabilities. In Stage 2, an EDKL meta-learner aggregates predictions from heterogeneous deep models (MLP, CNN, and LSTM) through kernel ridge regression with hybrid kernels optimized by a meta-heuristic search algorithm. This hybrid architecture supports robust, fine-grained forecasting across temporal, behavioral, and relational dimensions. Experiments on publicly available Twitter and MHASN datasets demonstrate substantial improvements over transformer-based and single-stage baselines, achieving up to 99% accuracy with consistently low error variance. The study also addresses ethical considerations related to privacy, bias, and potential misuse, emphasizes reproducibility through transparent experimental protocols, and outlines promising directions for future multimodal extensions, including richer linguistic, visual, and interaction signals for clinically relevant mental-health monitoring.