Answers are grounded in this paper and cite their sources.
Most digital mental-health tools give every user the same content, regardless of how that user feels now or what has helped them before. We built a pipeline that reads free text, detects stress and emotion with domain-adapted language models, and then picks a supportive activity with a contextual bandit. We also asked when per-user personalisation pays off. With five seeds per model, MentalRoBERTa outperformed retrained BERT-base and RoBERTa-base on stress detection (weighted F1 0.818). On our five-class, intervention-oriented emotion taxonomy it did not (macro F1 0.571). In a simulated engagement environment, simple non-contextual bandits did better than LinUCB at the population level. Per-user LinUCB gained nothing when it started from a shared population prior. Started from scratch, it beat the shared model by 0.179 in mean reward over 20 seeds, but only when simulated users had different activity preferences. We tested the classifiers on 70 free-text responses from real users, rated blind by a psychologist. Stress was detected well (area under the ROC curve 0.940). Emotion agreement was only fair (Cohen's κ 0.278), and most disagreements involved responses the psychologist labelled as overwhelm. The pipeline works technically, but its recommendation benefits come from simulation and need prospective testing.
No takes yet. Share an insight, caveat, or question.
Makaeva et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: