ADAPT-MIND: A Cross-Domain Deep Learning Framework for Mental Health Signal Detection in Shakespearean and Social Media Texts

Abstract

Cross-domain detection of mental distress signals in historical and contemporary texts is very challenging because of linguistic heterogeneity and the limited annotation resources. Adapter-Augmented Distillation And Prompt-Tuning – Memory-Augmented INference for Distress (ADAPT-MIND) proposes a unified framework that integrates adapter-enhanced LLM embeddings, graph-attentive temporal convolutions, memory-augmented transformers, adversarial–contrastive distillation and mutual information regularization, and mixed meta-learned soft prompts. Shared pre-processing and feature augmentation are applied to both Shakespearean utterances (10% expert-annotated, $\kappa = 0.82$) and 7489 student social-media posts (80/10/10 split), including pronoun variance, affective bigrams, syntactic entropy, and graph positional encodings. On five-fold cross-validation, ADAPT-MIND achieves accuracies of 0.912 for Shakespearean and 0.857 for students, with F1 scores of 0.911 and 0.856 and AUC of 0.948 and 0.903, respectively. Extensive ablation studies confirm an F1 drop of up to 4.2% by removing the graph-temporal module, proving its critical role in capturing multiscale narrative context. Error analysis shows misunderstanding of archaic idioms is responsible for 31%, while the rest is mainly due to mistaken academic stress, totaling 27%. Hyper-parameter optimization utilizes hierarchical Bayesian search over loss weights, learning rates, adapter dimensions, and convolution depth. The proposed approach delivers state-of-the-art performance across domains while ensuring interpretability and efficient parameter adaptation.