Specific topics, specific symptoms: linking the content of recurrent involuntary memories to mental health using computational text analysis

The study addresses whether recurrent involuntary autobiographical memories (IAMs) are maladaptive and how their characteristics relate to mental health. Despite being common in the general population, recurrent IAMs have been described as transdiagnostic features of disorders such as depression, anxiety, and PTSD. Prior work indicated that negative valence in recurrent IAMs correlates with higher psychopathology, but valence may be confounded with content. The research questions are: (1) Are specific content topics within recurrent IAMs associated with mental health symptoms beyond valence? (2) Is any such content disorder-specific across symptom domains? The study aims to disentangle content and valence and test their unique links with symptoms in a large nonclinical sample.

Past content analyses of autobiographical memories identified common themes (e.g., accidents, holidays, relationships). Some studies suggest IAM content varies by mental health status and may be disorder-congruent (e.g., health anxiety narratives focus on illness; social anxiety narratives show more self-referential and sensory language). However, mixed findings exist, with some studies finding no content differences by psychopathology or across disorders (e.g., social anxiety vs. panic disorder). Limitations of prior work include small samples, coarse single-category coding of content, and failure to disentangle valence from content. This motivated the use of large-scale, mixed-membership topic models and joint modeling of valence and symptoms.

Design and sample: Cross-sectional online survey administered in five waves between September 2018 and February 2020 to University of Waterloo undergraduates (N = 6187; mean age 19.9, SD 3.3; 71% women). Of these, 3624 participants who reported at least one recurrent IAM in the past year provided free-text descriptions of their most frequently recurring IAM. Measures: Recurrent Memory Scale assessed occurrence and characteristics (including valence from −2 very negative to +2 very positive). Mental health indices included DASS-21 (depression, anxiety, stress; focus on DASS-D), PCL-5 (PTSD), SPIN (social anxiety), and STICSA-T (trait anxiety). All measures showed high internal consistency. Data preparation: Supervised machine learning flagged and removed 202 invalid texts (71 human-labeled, 131 model-predicted). Text preprocessing followed best practices: tokenization, cleaning, stopword removal, vocabulary pruning, and lemmatization. A bag-of-words unigram representation was used. Topic modeling: Structural Topic Modeling (STM; R 'stm' package) discovered topics in IAM descriptions, with covariates for topic prevalence including valence and mental health symptom scores (DASS-D, PCL-5, SPIN, STICSA-T). Number of topics was selected via internal metrics and external human validation across models with 5–25 topics, yielding a final 16-topic solution. The final topic set included: Stressful events; Negative past relationships; Physical activities & performance; Embarrassing events; Close relationships; Illnesses, injuries, & deaths; Confrontations, fights, & arguments; Abuse & trauma; Conversations; Environments & locations; Interactions with friends; Communication & miscommunication; Subjective experiences of retrieval; Detailed/time-specific recollections; Experiences with family members; Reflections on decisions. Analysis: Regression of topic prevalence on symptom measures simultaneously (mutual adjustment) and valence, testing unique associations. Correlations between valence and symptoms were also computed. Sensitivity analyses excluded participants above clinical cutoffs.

• Negative valence of recurrent IAMs was significantly associated with greater symptoms across depression, PTSD, social anxiety, and general anxiety (all ps < .001; rs ≤ −.16), replicating prior findings. Effects persisted when excluding participants above clinical cutoffs (ps < .03; rs ≤ −.07).
• Symptoms uniquely predicted topic prevalence beyond valence and other symptoms: • Depression (DASS-D): positively associated with topic 8 “Abuse & trauma” (e.g., assault, abuse, trauma). Example coefficient: B = 0.0016, SE = 0.00060, p = .009. • PTSD (PCL-5): positively associated with topic 2 “Negative past relationships” (B = 0.00069, SE = 0.00025, p = .006); negatively associated with topic 4 “Embarrassing events” (B = −0.00044, SE = 0.00018, p = .02), topic 9 “Conversations” (B = −0.00042, SE = 0.00017, p = .01), and topic 11 “Interactions with friends” (B = −0.00037, SE = 0.00019, p = .049). • Social anxiety (SPIN): positively associated with topic 12 “Communication & miscommunication” (B = 0.00054, SE = 0.00026, p = .04) and topic 16 “Reflections on decisions” (B = 0.00027, SE = 0.00011, p = .02); negatively associated with topic 2 “Negative past relationships” (B = −0.00082, SE = 0.00026, p = .002) and topic 8 “Abuse & trauma” (B = −0.00039, SE = 0.00017, p = .03). • General anxiety (STICSA-T): positively associated with “Conversations” content (topic 9; as described in the results text) indicating more social/interpersonal conversational themes with higher trait anxiety.
• Several topics were seemingly benign (e.g., “Physical activities & performance,” “Environments & locations”) showing no consistent links with symptoms.
• No topic was universally related to all disorder symptoms; content associations were disorder-specific, whereas negative valence showed transdiagnostic associations.

Findings show that recurrent IAMs’ negative valence is transdiagnostically related to higher symptoms across depression, PTSD, social anxiety, and general anxiety, supporting models of IAMs as a process implicated in psychopathology. Critically, beyond valence, specific content topics uniquely map onto symptom domains, indicating disorder-specific content signatures. For example, abuse/trauma content aligns with higher depression symptoms; negative past relationship content aligns with PTSD symptoms, while positive/social topics are less expressed with higher PTSD, consistent with theories of reduced access to positive memories in PTSD. Social anxiety relates to themes of communication errors and decision reflections, consistent with disorder-specific social threat appraisals; general anxiety relates to conversational themes, possibly reflecting interpersonal worries common in generalized anxiety. Thus, content provides incremental, distinguishable information about mental health beyond emotional valence, addressing the core question of differentiating maladaptive versus benign recurrent IAMs.

This large-scale, computational text analysis demonstrates that both valence and content of recurrent IAMs relate to mental health. Negative valence is broadly (transdiagnostically) associated with higher symptom severity, while specific topics show disorder-specific associations, indicating that the nature of remembered events and how they are reconstructed matter beyond how negative they feel. The study presents a validated 16-topic structure for recurrent IAMs and quantifies content–symptom links in a large nonclinical sample. Future research should: compare involuntary versus voluntary autobiographical memories; examine other spontaneous thought types (e.g., future thinking, rumination, worry); sample multiple recurrent IAMs per person; use longitudinal/EMA designs to establish temporal dynamics; include clinical samples; and leverage NLP models that capture multiword expressions and word order to refine content detection.

• Focus on recurrent involuntary autobiographical memories limits generalizability to other memory types (e.g., voluntary AMs).
• Cross-sectional, single-session online design precludes causal inference; transient mood could influence both memory valence and symptom reports.
• Convenience sample of undergraduates (nonclinical) may not generalize to clinical populations; dimensional symptom approach may miss categorical differences.
• Only the most frequently recurring IAM was described; participants experience multiple recurrent IAMs, potentially limiting content diversity captured.
• NLP approach used unigrams and a bag-of-words representation, which omits multiword expressions and word order, potentially reducing interpretability and nuance of topics.
• Potential discrepancies between different analytic outputs (e.g., figures vs. tables) warrant cautious interpretation of some topic–symptom links.