hide
Free keywords:
-
Abstract:
Background: Suicide presents a formidable public health challenge, often eluding early detection due to the unreliability of self-report measures. Our work introduces a computational framework integrating behavioral, cognitive, and neural analyses to better understand suicide vulnerabilities and inform tailored interventions. Specifically, the model space we proposed explores how heightened inhibitory Pavlovian influence can disrupt instrumental control, leading to an active escape bias under stress.
Methods: We employed Reinforcement Learning and Active Inference models across two aversive Go/No-Go paradigms involving probabilistic cue-outcome learning. Two datasets were analyzed: Dataset 1 comprised N=150 patients with varying suicidality histories; Dataset 2 included N=55 suicidal patients and N=55 matched controls. In both tasks, participants needed to learn cue-outcome contingencies in order to minimize punishment (aversive sound) by responding or withholding a response. All models aimed to capture Pavlovian tendencies and recency effects in the data and were trained on each individual subject. Model efficacy was assessed via model evidence and exceedance probabilities (xp), alongside specific behavioral measures like Pavlovian biases.
Results: In both datasets, models effectively captured Pavlovian biases. Dataset 1's best model emphasized constant Pavlovian biases and a forgetting parameter (xp=0.99). Dataset 2's winning model highlighted constant Pavlovian biases without recency effects. Across datasets, significant differences in learning rate (lr, p < .05) and Go-to-Escape bias (we, p < 0.01) were noted between participants with and without suicidal thoughts and behaviors.
Conclusions: Reinforcement Learning models with constant Pavlovian biases best explain behaviour and performance biases in clinical groups using aversive Go/NoGo paradigms, highlighting the persistent nature of Pavlovian biases in psychiatric cohorts.
