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

The study addresses a longstanding problem in the neuroscience of deception: whether putative neural predictors actually measure deception-specific processes or instead reflect co-occurring, confounding processes such as arousal, risk–reward evaluation, belief inference, or self-interest. Despite advances in multivariate pattern analysis and economic signaling game paradigms that allow voluntary dishonest behavior, construct validity—especially discriminant validity—has been rarely tested. The authors’ central questions are: (1) Can a whole-brain fMRI-based predictor trained to distinguish deceptive from honest behavior truly index deception-specific neural processes? (2) If confounding signals are present, can they be identified and removed while retaining deception-specific predictive power? To investigate, the authors design isomorphic signaling games differing only in whether messages have a truth value (deception task) versus non-truth-apt preferences (control task). They test criterion validity (within-task prediction) and, crucially, discriminant validity by evaluating cross-task generalization to the control task that should not involve deception. They further develop a dual-goal tuning method to constrain predictors to perform at chance on the control task while maintaining predictive performance on deception, thereby improving construct validity.

The paper situates its contribution within over a century of skepticism toward lie detection methods, from early physiological measures (e.g., polygraph) to contemporary neuroimaging, where validity concerns persist (National Research Council, 2003). Recent cognitive neuroscience work has shifted from instructed-lie paradigms to motivated deception in signaling games, better capturing communicative deception (e.g., Gneezy 2005; Zhu et al. 2014; Jenkins et al. 2016). Multivariate pattern analysis and whole-brain decoders have shown promise for decoding mental states and predicting behavior (e.g., Wager et al. 2013; Chang et al. 2015; Speer et al. 2020). However, demonstrations of cross-task generalization often emphasize convergent validity, whereas discriminant validity—ensuring that predictors are not driven by related but distinct processes—has been under-examined. Meta-analyses implicate a network in deception (anterior insula, ACC, inferior frontal gyrus, inferior parietal lobule, superior frontal gyrus), but whether signals in these regions are specific to deception remains unclear (Farah et al. 2014; Lisofsky et al. 2014; Meier et al. 2021). Prior work has shown cases of separable neural representations (e.g., physical pain vs. social rejection; Woo et al. 2014) using generalization tests, yet methods for actively removing confounding signals across tasks are limited. The authors build on machine learning strategies (e.g., ridge, PLS, T-PLS) and concepts from constrained optimization and interpretable ML to develop a procedure that penalizes cross-task generalization to improve discriminant validity.

Participants: 40 healthy adults (16 males; mean age 20.8 ± 2.6 years) provided informed consent. Seven were excluded for one-sided choices (>90% of trials in at least one task), leaving N=33 for analysis. Procedures approved by Peking University IRB.

Tasks: Two isomorphic signaling games capturing motivated communication were used. On each trial, participants (senders) viewed two payoff allocations for themselves and a receiver; receivers (not present) were modeled as choosing the sender’s suggested option with 78% probability (no trial-by-trial feedback). One option favored the sender (selfish), the other favored the receiver (altruistic). In the deception task, messages had truth values (e.g., “Option A will earn you more money than B”), allowing truthful vs. deceptive signaling. In the control task, messages were non-truth-apt preferences (e.g., “I prefer that you choose option A”), enabling selfish vs. altruistic choices without deception. Bonus payment was determined by randomly selecting one trial from each task and carrying out the suggested option with 78% probability.

Behavioral measures: Proportions of selfish choices were compared across tasks; individual differences in deceptive (deception task) vs. selfish (control task) choice rates were examined.

fMRI preprocessing and activity estimation: BOLD activity was modeled with GLMs. Subject-level estimates: for each subject and run, one image for truthful/altruistic and one for deceptive/selfish choices (averaged across trials). Trial-level estimates: beta-series regression with a separate regressor per trial. Regressors: impulse at button press, convolved with double-gamma HRF. Nuisance regressors: average CSF signal, average white matter signal, top 10 aCompCor components (from fMRIPrep), and 24 motion parameters (6 affines at current and previous TR and their squares). Z-statistics of beta coefficients were used as MVPA inputs.

Prediction framework: Leave-one-subject-out cross-validation (LOSO). Whole-brain decoders were trained on deception-task data from 32 subjects using Thresholded Partial Least Squares (T-PLS) and tested on the held-out subject (for both deception and control tasks). Performance metrics: subject-level prediction accuracy for two-category images; trial-level AUC to handle class imbalance.

Discriminant validity test (overgeneralization): A deception-trained predictor was applied to control-task data. Significant correlation of predictor output with selfish vs. altruistic choices in the control task indicates shared (confounding) signals.

Confound-mitigation methods: Four approaches incorporated control-task data into training:

• Region removal: Mask voxels significantly correlated with control-task behavior at varying P-value cutoffs (e.g., P<0.05 up to P<0.99) and retrain.
• Relabeling (binary relevance): Combine both control-task categories as “truth” and deception-task lies as “lie” to encourage down-weighting of control-related signals.
• Regress-out: Leverage isomorphism (paired trials) to regress behavioral variation in control task out of deception-task signals (variants described in SI).
• Dual-goal tuning (constrained optimization): Introduce a penalty for cross-task covariance between model predictions and control-task behavior. Starting from a naïve map b, orthogonalize it with respect to the control-task brain–behavior covariance vector C2 using a Gram–Schmidt-like subtraction: b ← b − ω (C2 b)/(C2 C2), where ω controls the degree of orthogonalization. Conceptually, the method shears the predictor to enforce chance performance on the control task while preserving deception-task prediction.

Searchlight MVPA: Sphere radius 2 voxels (33 voxels) with PLS and LOSO CV to identify deception-predictive regions. Whole-brain permutation testing with TFCE correction (P<0.05). Overgeneralization assessed per searchlight by applying deception-trained local predictors to control task; proportion of voxels showing significant generalization computed across thresholds. Dual-goal tuning applied post hoc to searchlight maps to identify regions retaining deception prediction without control-task generalization.

Data and code: Dataset available on OpenNeuro (DOI: 110.18112/openneuro.ds005128.v1.0.0).

• Behavioral dissociation: Senders recommended selfish options less often in the deception task than in control (54.6% vs. 61.9%; paired t-test P=0.0074). Individual differences in deceptive vs. selfish choices were only weakly related (r=0.28, P=0.11), indicating partial dissociation.
• Criterion validity (within-task prediction): Whole-brain deception predictor significantly distinguished deceptive vs. truthful choices.
  • Subject-level: 78.8% ± 7.24% accuracy (mean ± SE), P<0.001.
  • Trial-level: AUC 56.6% ± 2.1%, P=0.004.
• Discriminant validity (overgeneralization to control task): Deception predictor outputs correlated with selfish choices in the nondeceptive control task.
  • Subject-level: r=0.39, P=0.021.
  • Trial-level: r=0.084, P=0.014.
  • Classification performance in control task was statistically indistinguishable from the deception task (subject-level: 69.7% vs. 78.8%, P=0.45; trial-level AUC: 55.4% vs. 56.6%, P=0.71), indicating compromised discriminant validity.
• Confound-mitigation comparisons:
  • Region removal, relabeling, and regress-out reduced overgeneralization but at the cost of reduced predictive performance on deception (tradeoff between accuracy and validity).
  • Dual-goal tuning nearly eliminated overgeneralization while preserving deception prediction:
    • Overgeneralization reduced to r=−0.0067 (P=0.85).
    • Trial-level deception AUC remained significant: 56.0%, P=0.01.
• High-confound test (distinguish deception vs. selfish nondeceptive trials): Only dual-goal tuning achieved above-chance discrimination (AUC=53.3%, P=0.0177) and outperformed each alternative method (P<0.05 for all pairwise comparisons).
• Neural localization (searchlight): Regions including superior frontal gyrus, precuneus, ACC, dlPFC, and NAcc predicted deception under naïve models (TFCE-corrected P<0.05). Overgeneralization at ROI level exceeded chance (permutation P=0.003). Depending on the significance threshold, 18–88% of voxels showed cross-task generalization (P<0.05: 18%; P<0.1: 28%; P<0.2: 39%; P<0.5: 66%; P<0.8: 88%). After dual-goal tuning, some regions (e.g., occipital pole) lost predictive power (fully confounded), whereas others (e.g., superior frontal gyrus) retained significant deception prediction, indicating recoverable deception-specific signals.

The work demonstrates that although a whole-brain fMRI predictor can distinguish deceptive from honest behavior, much of its predictive power can reflect confounding processes shared with nondeceptive selfish decision-making. This directly addresses the construct validity challenge, showing that naïve predictors may overgeneralize to related but distinct behaviors and therefore lack discriminant validity. By introducing a dual-goal tuning approach that penalizes cross-task covariance with a matched control task, the authors show that neural predictors can be reshaped to retain deception-related signals while eliminating reliance on nuisance signals. The positive high-confound test further corroborates improved discriminant validity by showing that the corrected predictor can distinguish deception from selfish nondeception, whereas alternative methods cannot. At the regional level, post-correction retention of prediction in higher-order networks (e.g., SFG, ACC) suggests that deception engages domain-general control and social-cognitive processes alongside deception-specific components, while predictive signals in unimodal sensory regions likely reflect confounds. These findings refine our understanding of the neural basis of deception and provide a methodological template for improving validity in neural prediction of complex constructs beyond deception, including applications in cognitive neuroscience and computational psychiatry where co-occurring processes are common.

This study identifies and quantifies confounding processes in fMRI-based deception prediction and introduces dual-goal tuning to improve discriminant validity by enforcing chance performance on a nondeceptive, isomorphic control task. The method preserves predictive power for deception while eliminating overgeneralization, and it uniquely enables above-chance discrimination between deceptive and merely selfish behavior. Regionally, it reveals that some areas’ predictive power was confounded, whereas others retain deception-specific information after correction. These contributions strengthen the scientific basis for studying the neurobiology of deception and offer a general strategy for isolating complex cognitive constructs. Future work should develop more naturalistic, open-ended deception paradigms, refine approaches to account for individual heterogeneity in out-of-sample prediction, and extend dual-goal tuning to other confounded domains (e.g., working memory vs. attention; valuation vs. salience; emotion dimensions) and clinical biomarker development.

• The signaling paradigm, while superior to instructed-lie tasks, still constrains how participants can deceive and may not capture the open-ended nature of real-world deception.
• Inferences about improved discriminant validity from reduced overgeneralization can rely on accepting the null; the authors mitigate this with a positive high-confound test, but statistical power limits remain possible.
• Individual heterogeneity in neural predictors and out-of-sample performance was not the focus and remains to be systematically characterized—especially important for forensic/clinical applications.
• Region-removal, relabeling, and regress-out methods showed tradeoffs between accuracy and validity; generalizability of dual-goal tuning across paradigms and modalities needs further evaluation.
• The isomorphic control task matches structure but differs in surface labels; differences in engagement of executive or social-cognitive processes by labels could still contribute to residual differences.
• The study’s moderate sample size (N=33 analyzed) may limit detection of smaller effects and finer-grained subgroup analyses.