Sensorimotor features of a reversal learning task bias decision behavior without disrupting individual difference structure

In computational psychiatry, task-irrelevant factors such as a task’s perceptual and motor features are typically assumed not to bias decision behavior, but at most to add noise. Yet growing evidence links sensorimotor processing to decision-making through multiple pathways, challenging this assumption. We tested this directly using a two-choice probabilistic reversal learning task completed by 90 participants under two sensorimotor conditions: a stationary condition requiring only arm movements to respond, and an active condition requiring participants to walk between physically separated computers. Task performance, measured as the propensity to choose the more probable rewarding option, did not differ between conditions. However, stay rate (the tendency to repeat the previous choice) was significantly elevated in the stationary condition, but only among participants who completed the active condition first; those who completed the stationary condition first showed no such difference. Logistic regression weights capturing the influence of recent win history on choice showed the same pattern, with stationary-condition elevation restricted to the active-first group. Cross-condition correlations for both measures were strong, indicating that this order-dependent effect shifted absolute measurements without disrupting individual difference structure. Prior sensorimotor history therefore does not compromise the task’s ability to extract stable cognitive variables, but does bias the values it yields — a non-random source of variance that, if it generalizes beyond this experiment, could shift patients across diagnostic thresholds in proposed clinical applications. The finding adds to a growing list of potential methodological contaminants in computational psychiatry’s cognitive-task paradigms that must be characterized and mitigated before clinical translation.

Improvements to AI Systems:

1. Order-Aware Bias Correction Module: Implement a trainable module in cognitive-assessment AI that detects and corrects for sensorimotor-order effects (e.g., first-task vs. second-task condition) by modeling prior task history as a latent variable, rather than assuming noise. This allows the AI to output bias-adjusted cognitive metrics (e.g., stay rate, win-history weights) that are invariant to task administration order.

2. Contextual State Embedding for Decision Models: Enhance reinforcement learning (RL) agents used in psychiatric diagnostics by adding a sensorimotor context encoder that ingests physical task conditions (e.g., walking vs. stationary) and their sequence. The AI can then disentangle stable cognitive traits (e.g., learning rate, choice propensity) from transient sensorimotor-induced shifts, improving cross-session reliability of patient stratification.

3. Adaptive Threshold Calibration for Clinical Triage: Build a calibration layer that, given a patient’s sensorimotor history (e.g., which condition was performed first), dynamically adjusts diagnostic thresholds for binary classifications (e.g., healthy vs. at-risk). This prevents order-dependent false positives/negatives, ensuring that a patient’s score is not misclassified due to task administration artifacts.

4. Counterfactual Simulation for Experimental Design: Use the paper’s finding to train a generative model that simulates how different sensorimotor condition orders would alter a given participant’s cognitive metrics. This allows AI-driven experimental design tools to pre-test task protocols and recommend optimal condition ordering to minimize bias before human data collection.

5. Robust Individual-Difference Preservation: Implement a normalization algorithm that leverages the strong cross-condition correlations observed (e.g., stay rate) to project raw scores onto a common latent space, removing order-dependent absolute shifts while preserving rank-order of individual differences. This improves AI’s ability to compare patients across studies with different task protocols.

What the Improved AI System Can Do:

• Deliver bias-free cognitive scores for psychiatric assessments, even when patients complete tasks under varying physical conditions or in different orders.

• Predict and flag when a patient’s diagnostic score may be artificially inflated/deflated due to sensorimotor history, prompting a re-test or adjusted interpretation.

• Automatically design clinical trial protocols that minimize order effects, reducing methodological variance and increasing statistical power.

• Provide personalized task administration—e.g., recommend a specific condition order for a patient to maximize measurement stability based on their prior exposure.

• Enable meta-analyses across heterogeneous studies by harmonizing cognitive metrics from different sensorimotor contexts, improving AI’s ability to learn from aggregated clinical data.

In computational psychiatry, task-irrelevant factors such as a task's perceptual and motor features are typically assumed not to bias decision behavior, but at most to add noise. Yet growing evidence links sensorimotor processing to decision-making through multiple pathways, challenging this assumption. We tested this directly using a two-choice probabilistic reversal learning task completed by 90 participants under two sensorimotor conditions: a stationary condition requiring only arm movements to respond, and an active condition requiring participants to walk between physically separated computers. Task performance, measured as the propensity to choose the more probable rewarding option, did not differ between conditions. However, stay rate (the tendency to repeat the previous choice) was significantly elevated in the stationary condition, but only among participants who completed the active condition first; those who completed the stationary condition first showed no such difference. Logistic regression weights capturing the influence of recent win history on choice showed the same pattern, with stationary-condition elevation restricted to the active-first group. Cross-condition correlations for both measures were strong, indicating that this order-dependent effect shifted absolute measurements without disrupting individual difference structure. Prior sensorimotor history therefore does not compromise the task's ability to extract stable cognitive variables, but does bias the values it yields - a non-random source of variance that, if it generalizes beyond this experiment, could shift patients across diagnostic thresholds in proposed clinical applications. The finding adds to a growing list of potential methodological contaminants in computational psychiatry's cognitive-task paradigms that must be characterized and mitigated before clinical translation.