Screening anxiety via contrastive autobiographical recall

Language offers a low-burden and scalable pathway for digital anxiety screening, particularly in telehealth or repeated-monitoring settings where spontaneous speech may already be available. This study introduces a contrastive autobiographical recall framework that uses short positive and negative personal memories to capture within person affective shifts in language. By modelling how the same individual expresses emotionally distinct experiences, the proposed approach aims to identify anxiety-related linguistic patterns that may not be captured from a single static text representation.

A total of 156 participants completed a 5-7 minute spontaneous speech task involving positive and negative autobiographical memories. Anxiety status was defined using HAM-A scores, yielding non-anxious (n = 101) and anxious (n = 55) groups. Transcripts were segmented using Qwen-2.5-7B-Instruct as a deterministic constrained extractor, preserving only verbatim positive and negative spans alongside the complete transcript. Positive, negative, and complete narratives were encoded with frozen BERT model and combined with a contrast vector capturing within-person affective shift. Performance was evaluated using a leakage-safe leave-one-out cross-validation pipeline.

The proposed pipeline achieved 70% accuracy and 0.67 macro-F1 across leaveone-out folds, with stronger performance for non-anxious participants than anxious participants. Bootstrap confidence intervals were 0.62-0.77 for accuracy and 0.59-0.75 for macro-F1. Ablation analysis showed that the full composite representation provided the best balanced performance and strongest anxious-class detection. The method also outperformed BERT-based and lexicon-based baseline models.

These findings suggest that short autobiographical speech can provide a useful complementary signal for digital anxiety screening when modelled with contextual embeddings and within-person affective contrast. Latent-space augmentation supported learning in this small cohort without altering participant-authored language. However, anxious-class sensitivity was moderate, and HAM-A labels should be interpreted as screening rather than diagnostic labels. Further validation in larger and more diverse clinical cohorts is needed.