Joint longitudinal and survival modeling predicts avoidance decisions

Poster Session D - Monday, April 15, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Brooke Staveland¹ (bstavel@berkeley.edu), Julia Oberschulte², Olivia Kim-McManus^3,4, Jon Willie^5,6, Peter Bunner^5,6, Mohammad Dastjerdi⁷, Jack Lin^8,9, Ming Hsu^1,10, Robert Knight^1,11; ¹UC Berkeley, ²Department of Psychology, LMU Munich, ³Division of Neurology, Rady Children’s Hospital, ⁴Department of Neurosciences, UC San Diego, ⁵Department of Neurosurgery, Washington University School of Medicine, St. Louis, ⁶National Center for Adaptive Neurotechnologies, ⁷Department of Neurology, Loma Linda University, ⁸Department of Neurology, UC Davis, ⁹Center for Mind and Brain, UC Davis, ¹⁰Haas School of Business, UC Berkeley, ¹¹Department of Psychology, UC Berkeley

Decision making requires approaching and avoiding stimuli representing rewarding and aversive outcomes. Approach-avoidance is dependent on the temporal dynamics of the decision. Both when an actor chooses to approach/avoid and what environmental changes occur are crucial to adaptive decisions. Modeling temporal dynamics of continuous predictors (e.g., increasing threat or diminishing reward) leading to a single choice (switching from approach to avoidance), is difficult. We apply models that jointly account for dependencies between survival outcomes (when) and longitudinal measurements (what), to an approach-avoidance conflict task (Pacman). Decisions to move along a corridor involved potential gains (dots, resulting in points) and losses (ghost attack, resulting in death). Our joint models build on a linear mixed effects model, capturing how a predictor (e.g., threat, reward) changes over time, combined with a time-to-event model, capturing event occurrence. These sub-models are linked via shared random effects structures using Bayesian model fitting. These models are effective at predicting trial-level subject behavior in an online sample (n=191), and in presurgical epilepsy patients (n=15). These models predict temporal avoidance decisions in held-out trials (average online AUC: 0.76 [95% CI:0.75-0.78]; average patient AUC: 0.80 [95% CI:0.76-0.85]). They outperform permuted models where predictors were shuffled (mean online AUC difference: 0.19 [95% CI:0.16-0.22]; mean patient AUC difference: 0.24 [95% CI: 0.15-0.32]). Additionally, we find preliminary evidence that these models can incorporate neural measures, such as hippocampal theta or prefrontal high-frequency activity, to predict avoidance decisions.

Topic Area: THINKING: Decision making