Distinguishing deception from its confounds by improving the validity of fMRI-based neural prediction

The study addresses persistent skepticism surrounding biologically based deception detection, which stems from difficulty in separating deception-specific neural processes from co-occurring factors such as arousal, self-interest, risk–reward evaluation, and belief inference. Leveraging advances in multivariate pattern analysis (MVPA) and economic signaling games, the authors ask whether neural predictors trained to distinguish deceptive from honest behavior truly capture deception-specific processes (construct validity), with emphasis on discriminant validity. They design an isomorphic control game that matches the deception task in players, strategies, and payoffs but removes truth-value from messages, enabling tests of overgeneralization: if a deception predictor also predicts selfish choices in the control task, it likely relies on signals shared across tasks rather than deception per se. The goals are to (a) quantify predictive accuracy for deception (criterion validity), (b) test for confounding via cross-task generalization (discriminant validity), and (c) develop methods, notably dual-goal tuning, to purge confounds while retaining deception-related signals.

The paper situates its contributions in a century-long discourse on deception detection, from early polygraph critiques to modern neuroimaging. Prior work has identified frontal and insular regions linked to deception, but concerns about reverse inference and confounds remain (NRC 2003; Farah et al., 2014). MVPA has enabled decoding of mental states across domains (e.g., pain, affect), and signaling games model motivated communication and deception in economics and biology. Meta-analyses implicate lateral/medial PFC and anterior insula in deception; more recent work shows cognitive control relates to honesty. Generalization tests have been used to assess construct distinctiveness, but methods to remove existing cross-task generalization are underdeveloped. The authors build on machine learning and constrained optimization ideas to formalize and remediate confounding in neural predictors.

Participants: 40 healthy adults (16 males; mean age 20.8 ± 2.6 years) consented; 7 excluded for >90% one-sided choices in at least one task; final N=33. IRB-approved at Peking University.

Tasks: Two isomorphic signaling games with identical players, strategies, and payoffs. On each trial, senders observed two payoff allocations (one favoring sender; one favoring receiver). Receivers could not see options and relied on the sender’s message. Deception task: messages had truth-value (e.g., “Option A will earn you more money than B”); senders chose truthful vs deceptive messages. Control task: messages lacked truth-value (e.g., “I prefer that you choose option A”); senders chose selfish vs altruistic recommendations without deception. No trial-by-trial feedback; receivers typically accept suggestions ~78% as informed; bonus payments determined by randomly selecting one trial per task and implementing suggestion with 78% probability.

Behavioral analyses: Compared proportions of selfish recommendations across tasks; assessed individual differences correlation between deception rate (deception task) and selfish rate (control task).

fMRI acquisition and decoding: Whole-brain MVPA to predict deception. Two prediction levels:

• Subject-level: For each subject, estimated one image for truthful/altruistic and one for deceptive/selfish choices by averaging trials of same type via GLM.
• Trial-level: Estimated a separate image per trial using beta-series GLM.

GLM details: Regressors time-locked to button press, convolved with double-gamma HRF. Nuisance regressors: average CSF and white matter signals, top 10 PCA components from CSF+WM masks (a_comp_cor), and 24 motion parameters (6 affine transforms for current and previous TR, plus their squares). Z-statistics of coefficients used for decoding inputs.

Prediction framework: Leave-one-subject-out cross-validation. Train a whole-brain predictor on deception task data from 32 subjects using Thresholded Partial Least Squares (T-PLS), then test on the held-out subject’s deception and control task data.

Discriminant validity test: Apply the deception-trained predictor to control task data to assess nuisance correlation (overgeneralization) with selfish vs altruistic choices at subject and trial levels.

Methods to control confounds:

1. Region-removal: Mask voxels significantly correlated with control-task behavior at varying P-value cutoffs.
2. Relabeling (binary relevance): Treat all control-task trials as “truth,” retaining “lie” labels only from deception task, to down-weight control-task signals.
3. Regress-out: Use isomorphism to pair trials across tasks and regress control-task behavior out of deception-task signals (variants detailed in SI).
4. Dual-goal tuning: Constrained optimization to penalize cross-task covariance between predictor scores and control-task behavior. Formal test of shared signal: cov(X₂b, Y₂) = bᵀC₂/n > 0. Extended cost function adds ω bᵀC₂. Implemented via two-step Gram–Schmidt orthogonalization: construct naïve map b, then orthogonalize with respect to C₂ using b − ω (C₂·b)/(C₂·C₂) C₂. Hyperparameters chosen to maximize deception-task performance while enforcing null generalization to control-task.

Searchlight MVPA: Whole-brain spherical ROIs (radius 2 voxels; 33 voxels) using PLS with leave-one-subject-out CV. Whole-brain permutation testing with TFCE to identify regions predicting deception (P<0.05). ROI-level cross-task generalization quantified; dual-goal tuning applied per searchlight to isolate deception-specific signals.

• Behavioral dissociation: Senders recommended selfish options less often in the deception task than in the control task (54.6% vs 61.9%; paired t-test P=0.0074). Individual differences in deception vs selfishness weakly related (r=0.28, P=0.11; R²=0.08).
• Subject-level prediction: Whole-brain neural predictor distinguished deceptive vs truthful average images significantly above chance (78.8% ± 7.24% SE; P<0.001). Confusion rates: TP 39.4%, TN 39.4%, FP 10.6%, FN 10.6%.
• Trial-level prediction: Significant discrimination at the single-trial level (AUC = 56.6% ± 2.1% SE; P=0.004).
• Discriminant validity failure (overgeneralization): Deception-trained predictor correlated with selfish (nondeceptive) choices in control task—subject-level r=0.39 (P=0.021); trial-level r=0.084 (P=0.014). Classification performance in control task statistically indistinguishable from deception task (subject-level accuracy 69.7% vs 78.8%; P=0.45; trial-level AUC 55.4% vs 56.6%; P=0.71).
• Confound-removal methods comparison: Region-removal, relabeling, and regress-out reduced nuisance correlation but at the cost of predictive accuracy, often substantially, and with incomplete orthogonality.
• Dual-goal tuning: Nearly eliminated overgeneralization (nuisance correlation r = −0.0067, P=0.85; reduction vs naïve P<0.001) while retaining deception prediction (single-trial AUC = 56.0%; P=0.01). Provided most favorable trade-off between accuracy and discriminant validity among methods tested.
• High-confound test (deceptive vs selfish trials): Only dual-goal tuning significantly distinguished deceptive from selfish nondeceptive choices (mean per-subject AUC = 53.3%; P=0.0177), outperforming other methods (P<0.05 vs each alternative).
• Searchlight analyses: Regions including superior frontal gyrus (SFG) and precuneus predict deception (TFCE P<0.05). Overgeneralization at ROI level more prevalent than chance (permutation P=0.003). Depending on cutoff for defining meaningful generalization, 18–88% of voxels overgeneralized (P<0.05: 18%; P<0.1: 28%; P<0.2: 39%; P<0.5: 66%; P<0.8: 88%). Dual-goal tuning revealed heterogeneity: some regions (e.g., left occipital pole) lost predictive power after correction, implying confound-driven signals; others (e.g., SFG, ACC) retained deception-specific predictive signals after orthogonalization.

The findings show that naïve neural predictors trained to classify deception can achieve above-chance accuracy yet rely substantially on signals shared with nondeceptive, selfish decision-making, undermining discriminant validity. Cross-task generalization provides a powerful test of construct validity, complementing within-task sensitivity/specificity (criterion validity). By introducing a constrained, dual-goal tuning procedure that actively suppresses cross-task covariance with control-task behavior, the authors isolate deception-related signals while maintaining predictive performance. This approach clarifies the neural systems implicated in deception, indicating that predictive signals in some sensory/unimodal areas are confound-driven, while polymodal, higher-order regions (e.g., SFG, ACC, dlPFC, NAcc) retain deception-specific information consistent with hypotheses regarding executive control, self-referential processing, and social cognition. The positive high-confound test further strengthens evidence for improved discriminant validity, demonstrating the corrected predictor can distinguish deception from merely selfish choices. More broadly, the methodology offers a framework to disentangle co-occurring cognitive processes post hoc, informing both basic cognitive neuroscience and applied domains (e.g., forensic contexts, computational psychiatry) where specificity of neural biomarkers is paramount.

This work establishes a rigorous framework for assessing and improving the validity of fMRI-based predictors of deception. It demonstrates that naïve predictors overgeneralize to nondeceptive selfish choices, compromising discriminant validity, and introduces dual-goal tuning to suppress confounding signals while preserving deception-related predictive power. Searchlight analyses identify regions where deception signals can be isolated from shared processes, supporting models in which deception engages domain-general executive functions and social cognition alongside potentially deception-specific mechanisms. Future research should develop paradigms capturing the open-ended nature of real-world deception (when, how, whom to deceive), systematically manipulate engagement of executive and social processes across control conditions, and address individual heterogeneity to enhance out-of-sample prediction. The dual-goal approach may generalize to other complex constructs (e.g., working memory vs attention, valuation vs salience, emotion vs arousal) to refine neural biomarkers with improved construct specificity.

• Task constraints: Although participants chose when to lie, the paradigm still specified how to lie; ecological validity for open-ended deception is limited.
• Null-based inference risk: Discriminant validity tests that rely on absence of overgeneralization can suffer from low power; mitigated here with a positive high-confound test but still a general concern.
• Commingled signals: Interest and nuisance signals likely comingle at voxel/ROI levels, challenging region-removal approaches and complicating interpretations of localized effects.
• Generalizability and heterogeneity: Individual differences in deception and selfishness were modestly related; more work is needed to model inter-individual heterogeneity for clinical/forensic applications.
• Task isomorphism vs psychological labels: Despite isomorphic game structures, labels (truth vs preference) may differentially engage cognitive processes; thus non-generalization does not imply uniqueness to deception.
• Performance trade-offs: Methods that remove confounds often reduce predictive accuracy; even dual-goal tuning involves balancing signal suppression and retained performance.
• Lack of trial feedback and receiver behavior measurement in real time may limit inference about interactive dynamics of deception.