Abstract Sat906: Using Natural Language Processing to Distinguish Recalled Experiences of Death from Drug-Induced Hallucinations and Dreams in Cardiac Arrest and Critical Ill Survivors

Background: Around, 10% of cardiac arrest survivors report vivid Recalled Experiences of Death (RED)— heightened lucidity, visual and auditory awareness, and a perception of a purposeful life review - which positively impact the quality of survivorship. However, these experiences are often labelled as hallucinations or dream like experiences. As Natural Language Processing (NLP) - a scalable and reproducible method that enables systematic analysis of unstructured, self-reported data, allowing pattern recognition beyond subjective interpretation - we sought to determine whether RED may be objectively distinguished from dreams, and drug-induced states. Hypothesis: We hypothesized that NLP can objectively differentiate RED in cardiac arrest patients from other dream-like or hallucinatory states based on thematic content, offering insight into consciousness during clinical death. Methods: We analyzed anonymized first-person narratives from three publicly available databases: RED accounts from the Near-Death Experience Research Foundation (NDERF), dream experiences from DreamBank, and drug-induced reports from Erowid. RED cases were identified via keyword filtering using resuscitation-related terms (e.g., “CPR,” “defibrillator,” “chest compressions”). A Longformer-based transformer model was fine-tuned to classify entire narratives into one of three categories: RED, Dream, or Drug. A separate BERT-based model was trained to label individual RED sentences by predefined experiential themes (e.g., separation from the body, life review, return to life) or as non-relevant (“Other Experience”). Manual expert labeling, stratified sampling, and data augmentation enhanced class balance and model performance. Results: We analyzed 3,700 anonymized narratives: 1,245 RED, 1,190 dreams, and 1,265 drug-induced experiences. The Longformer model achieved a validation F1-score of 98% and 100% accuracy on the holdout dataset. It correctly identified all holdout drug narratives—even without substance names—demonstrating strong contextual generalization. The BERT-based model achieved a validation F1-score of 90% and a holdout F1-score of 87% in identifying RED-specific themes. Conclusion: Transformer-based NLP models can differentiate RED from other states and reveal their thematic patterns, suggesting these experiences are distinct and structured and unlike hallucinations. This provides method for analyzing survivor narratives and exploring psychological outcomes of cardiac arrest.