Cardiac Interoceptive Prediction in Humans and Large Language Models: Behavioral, ECG, and Simulated Data from a Heartbeat Expectation-Violation Task

Recorded variables include heartbeat predictions, prediction confidence, subjective heartbeat sensations, anxiety ratings, response times, arithmetic responses, and calculation accuracy. ECG signals were collected concurrently, allowing the derivation of heart rate, inter-beat interval, R-R interval, and root mean square of successive differences (RMSSD).The LLM component contains 1,000 independently generated simulated participants constructed through a role-playing paradigm.

Gemma-3-27B-IT was used as the primary model, while Qwen-3-32B-AWQ and GLM-4-32B were included in supplementary experiments. Each simulated participant was identified by a unique participant ID and random seed and was assigned demographic, personality, interoceptive, and affective characteristics. These included age, gender, education, occupation, sleep duration, current mood, anxiety, worry tendency, interoceptive awareness, bodily vigilance, baseline arousal, and Big Five personality traits.Each simulated participant completed a text-based version of the same 90-trial task.

Trial-level variables include cue identity, probability environment, task outcome, heartbeat prediction, task prediction, confidence, simulated cardiac sensation, anxiety rating, arithmetic performance, and quality-control fields such as formatting errors, role leakage, invalid-response reasons, and retry counts. These simulated reports were generated through textual reasoning and do not contain genuine physiological signals.