Schedule of Events | Search Abstracts | Invited Symposia | Symposia | Rising Stars | Poster Sessions | Data Blitz

Sketchpad Series 

Can Machine Learning Predict Naturalistic Episodic Memory Performance?

Poster Session A - Saturday, March 7, 3:00 – 5:00 pm, Fairview/Kitsilano Ballrooms

Lexi Golestani¹ (lexi.golestani@utah.edu), Tyler Friedholm¹, Cory Inman¹; ¹University of Utah

Episodic autobiographical memory performance varies widely across individuals, especially when recalling stressful experiences. Consistent with theories of expertise-driven schema formation, prior analyses of this dataset suggest that experience moderates memory performance under stress. However, other individual difference measures likely contribute to the complexity of this relationship. In this work-in-progress project, we employ supervised machine-learning models to investigate whether measures such as personality, risk propensity, stress mindset, subjective stress appraisal, and activity-based identity can predict episodic richness in a naturalistic autobiographical memory dataset. Participants (N ≈ 70) provided written narratives describing “stressful, intense, or demanding” and “non-stressful” whitewater paddling events (≈420 memories total). Narratives were automatically scored for episodic (“internal”) and semantic (“external”) detail using a validated transformer-based classifier (van Genugten & Schacter, 2024). We are currently developing and benchmarking predictive models of episodic detail using linear (ElasticNet) and tree-based approaches (Random Forest, XGBoost), with grouped train/test/validation splits to prevent overfitting. Planned analyses include permutation- or SHAP-based feature-importance mapping, as well as exploratory theory-driven feature engineering to test whether derived variables (e.g., identity × personality) improve model performance. The expected outcome is a ranked account of which psychological and contextual features most strongly predict episodic memory performance, and an evaluation of whether such prediction is feasible in this dataset at all. This project aims to demonstrate a scalable computational framework for linking individual differences to autobiographical memory, and illustrate how machine-learning tools can complement cognitive-theory–driven models of stress, self, and memory. Preliminary model performance comparisons and feature-importance visualizations will be presented.

Topic Area: LONG-TERM MEMORY: Episodic