0123 Large Language Model-Derived Dream Affect is a Prospective Predictor of PTSD Symptoms in Hospitalized Trauma Patients
Holland Doise, A. Gonzalez Quesada, Garrett Baber, Nancy Hamilton, Matthew Gratton, Anthony N Reffi, Lily Jankowiak, Gregory Mahr, David A. Moore, Christopher L Drake
Sleep May 1, 2026 DOI: 10.1093/sleep/zsag091.0123 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Prospective cohort study Peer reviewed |
|---|---|
| Sample size | 87 |
| Population | Patients recruited from a Level I trauma center in Detroit, MI, following traumatic injury |
| Duration | 2-month follow-up |
| Topics | Anxiety Dreaming PTSD |
| Keywords | Dream Affect linguistics Arousal Prospective cohort study Posttraumatic stress Injury prevention Poison control Insomnia Clinical psychology Severity of illness Suicide prevention Mental health |
| Key findings | LLM-derived negative affect and arousal from dreams reported during hospitalization predicted PTSD symptom severity one month later, while self-reported dream affect did not. |
Abstract
Abstract Introduction Nightmares are associated with more severe posttraumatic stress disorder (PTSD). The advent of large language models (LLM) offers unique advantages to efficiently code affective dream content over human methods, which can be resource intensive and prone to error. This study tested whether dream affect (DA), measured via self-report, human-rated narrative coding, and LLM scoring, prospectively predicted PTSD symptom severity following traumatic injury.
Methods: Patients (N = 87) recruited from a Level I trauma center in Detroit, MI, following traumatic injury completed surveys during hospitalization (T1), and 1 month (T2), and 2 months (T3) posttrauma. At each timepoint, participants provided a dream report, self-reported DA, and completed measures of PTSD symptom severity, fear of sleep, and insomnia severity. Dream narratives were scored by trained raters using the Disturbing Dream Content Inventory (DDCI) and by an LLM (Gemma 3 [12B]) to estimate negative affect (NA) and arousal. Using bootstrap-estimated regressions, we tested whether affect from dreams reported during hospitalization predicted PTSD severity at 1 and 2 months posttrauma, and whether DA at 1 month predicted PTSD severity at 2 months, adjusting for age, sex, and baseline PTSD, fear of sleep, and insomnia symptoms.
Results: T1 DA showed minimal predictive value for T3 PTSD symptoms (all ps > .12). However, higher Gemma-derived NA at T1 predicted higher PTSD symptoms at T2 (β = 6.03, p = .022), and higher T1 Gemma-derived arousal similarly predicted T2 symptoms (β = 9.86, p = .035). At T2, both Gemma-derived NA (β = 4.63, p = .038) and DDCI Total scores (β = 8.28, p = .010) predicted higher PTSD symptoms at T3. Self-reported DA showed no significant predictive associations.
Conclusion: LLM-derived affect from posttrauma dreams reported during hospitalization forecasted PTSD symptoms one-month posttrauma over and above other psychological risk factors. Interestingly, self-reported and human-coded DA from the hospitalized dreams did not predict later PTSD symptoms. However, one-month posttrauma DA measured by an LLM and by human coders predicted PTSD severity at two-months posttrauma. These findings suggest that dream narratives may carry clinically relevant emotional signals detectable by an LLM and trained human coders. Support (if any)