Machine learning approaches to uncover the neural mechanisms of motivated behaviour: from ADHD to individual differences in effort and reward sensitivity

Nam Trinh

q-bio.NCApr 13, 2026

Machine learning approaches to uncover the neural mechanisms of motivated behaviour: from ADHD to individual differences in effort and reward sensitivity

Authors: Nam Trinh

Organizations: School of Computing, Dublin City University

Abstract

Motivated behaviour relies on the brain's capacity to evaluate effort and reward. Dysregulation within these processes contributes to a spectrum of conditions, from hyperactivity in attention-deficit/hyperactivity disorder (ADHD) to diminished goal-directed behaviour in apathy. This thesis investigates the neural mechanisms underlying ADHD using electroencephalography (EEG) and examines individual differences in effort and reward sensitivity using neuroimaging, applying machine learning approaches through three main studies. In Study 1, task-based and resting-state EEG were employed with machine learning models to classify adult individuals with ADHD and healthy controls. Machine learning classifiers trained on task-based EEG during a stop signal task outperformed those trained on resting-state EEG, with the strongest predictive features arising from gamma-band spectral power over fronto-central and parietal regions. In Study 2, diffusion MRI and whole-brain permutation-based analyses identified associations between white matter integrity and computationally modelled parameters reflecting effort and reward sensitivity, with SMA-connected tracts emerging as a central hub. In Study 3, grey matter volumes from structural T1-weighted MRI were used to examine correlates of effort sensitivity, reward sensitivity, and subclinical apathy, with machine learning confirming robust decoding of reward sensitivity and apathy levels. Across studies, fronto-parietal circuits emerged as central to effort valuation and reward processing. These findings may serve as neural biomarkers for improving diagnostic accuracy in ADHD and motivational impairments, and for guiding personalised neurotechnological interventions.

Figures & tables

Figure 3.1 : Experimental design, EEG data acquisition, preprocessing, feature extraction and classification pipeline. A. EEG was recorded during two conditions: resting state with eyes closed and performance of the stop-signal task. B. In the stop-signal task, participants responded to directional arrows on “Go” trials and withheld responses when a stop signal appeared. C. EEG data were preprocessed with bandpass (0–95 Hz) and notch (48–52 Hz) filters, followed by common average re-referencing. Spectral power features were extracted from 2-s resting-state epochs and 1-s task-related epochs. D. Electrode placement followed the international 10–20 system. E. Spectral power features were used as inputs to machine-learning classifiers to distinguish ADHD from healthy control participants. Model training employed nested cross-validation, with outer folds for model validation and inner folds for hyperparameter tuning via grid search.

Figure 3.2 : Statistical analysis of classification performance with resting-state and stop-signal task EEG. A. Left, distribution of AUC across models compared with chance. Middle, permutation test with 1,000 iterations showed no significant deviation from chance (Monte Carlo $p=0.064$ ). Right, true-label performance was significantly greater than chance ( $t=124.99$ , $p<0.001$ ) and higher than permuted-label performance ( $t=42.36$ , $p<0.001$ ), whereas permuted-label performance did not differ from chance ( $t=1.32$ , $p=0.19$ ). B. Left, distribution of AUC across models compared with chance. Middle, permutation test with 1,000 iterations indicated significant deviation from chance (Monte Carlo $p=0.010$ ). Right, true-label performance was significantly greater than chance ( $t=164.5$ , $p<0.001$ ) and higher than permuted-label performance ( $t=67.98$ , $p<0.001$ ), while permuted-label performance did not differ from chance ( $t=-1.41$ , $p=0.14$ ). C. All classifiers trained on stop-signal task EEG achieved significantly higher AUCs than those trained on resting-state EEG (all FDR-corrected $p<0.001$ )

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.534 ± 0.035	0.626 ± 0.018	0.012 ± 0.024	0.024 ± 0.049	0.013 ± 0.025
SVM	0.462 ± 0.041	0.623 ± 0.020	0.015 ± 0.026	0.032 ± 0.054	0.019 ± 0.029
RF	0.519 ± 0.037	0.590 ± 0.028	0.176 ± 0.048	0.364 ± 0.106	0.223 ± 0.057
GB	0.494 ± 0.046	0.583 ± 0.033	0.189 ± 0.062	0.346 ± 0.115	0.226 ± 0.064
ET	0.530 ± 0.035	0.600 ± 0.024	0.123 ± 0.047	0.306 ± 0.127	0.164 ± 0.059
AdaB	0.536 ± 0.042	0.593 ± 0.034	0.271 ± 0.065	0.413 ± 0.095	0.308 ± 0.065

Table 3.1: Classification performance for ADHD versus healthy controls using resting-state EEG. Performance of nine classifiers: logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting (GB), extremely randomized trees (ET), AdaBoost (AdaB), decision tree (DT),

k

-nearest neighbors (KNN), and Gaussian naïve Bayes (GNB). Metrics are mean

\pm

s.d. across cross-validation folds: area under the ROC curve (AUC), accuracy, recall (sensitivity), precision (positive predictive value) and F1 (harmonic mean of precision and recall). ADHD was treated as the positive class.

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.691 ± 0.057	0.833 ± 0.017	0.971 ± 0.023	0.853 ± 0.011	0.907 ± 0.011
SVM	0.477 ± 0.091	0.833 ± 0.015	0.980 ± 0.022	0.847 ± 0.008	0.908 ± 0.010
RF	0.640 ± 0.045	0.833 ± 0.013	0.969 ± 0.012	0.854 ± 0.008	0.907 ± 0.008
GB	0.672 ± 0.062	0.835 ± 0.016	0.980 ± 0.021	0.849 ± 0.009	0.909 ± 0.010
ET	0.652 ± 0.046	0.830 ± 0.011	0.976 ± 0.010	0.846 ± 0.007	0.906 ± 0.006
AdaB	0.750 ± 0.048	0.837 ± 0.020	0.960 ± 0.019	0.864 ± 0.012	0.908 ± 0.012

Table 3.2: Classification performance for ADHD versus healthy controls using stop signal EEG data. Performance of nine classifiers: logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting (GB), extremely randomized trees (ET), AdaBoost (AdaB), decision tree (DT),

k

-nearest neighbors (KNN), and Gaussian naïve Bayes (GNB). Metrics are mean

\pm

Figure 3.3 : Top most important EEG features for machine learning evaluated with resting-state and stop-signal task EEG data. Permutation-based feature importance analysis revealed distinct spectral–spatial profiles across paradigms. A. In the resting-state condition, the highest-ranking predictors were theta/beta ratio (TBR) measures over parietal and central electrodes, with additional contributions from occipito-parietal beta- and alpha-band power. B. During the stop-signal task, model performance was driven primarily by relative gamma-band (45–49 Hz) power over fronto-central, temporal, and parietal sites, with secondary contributions from beta- and alpha-band activity. C. Group-level comparisons (independent-samples t-tests, FDR-corrected) of the top five resting-state features revealed a significant increase in C4 TBR in ADHD relative to controls, while parietal TBR features (P7, P5, P9) showed consistent but nonsignificant trends toward higher values in ADHD. D. For the stop-signal task, FDR-corrected tests demonstrated significantly reduced relative gamma-band power in ADHD at FC1, T8, and PO3, together with elevated relative mid-beta power at C6, whereas CP5 did not differ reliably between groups.

Figure 4.1 : Experimental protocol and data analyses. A. Diffusion MRI (dMRI) data acquisition. dMRI data were acquired to assess white matter microstructure, including FA and MD. B. Effort-based decision-making task and computational modelling of behaviour. Participants completed an effort-based decision-making task, choosing whether to perform biceps contractions of varying effort levels (20%, 40%, 60%, or 80% of maximal voluntary contraction) to obtain varying monetary rewards (1, 5, 10, or 20 euro cents). Computational modelling of their acceptance rates in this task, yielded individual effort and reward sensitivity parameters ( $\beta_{\text{Effort}}$ and $\beta_{\text{Reward}}$ , respectively). C. Whole-brain dMRI data analysis. dMRI data were exploited to compute voxel-wise FA and MD quantification in each subject (top left panel). We then applied a cluster-based analysis to identify significant clusters where FA and MD covaried with $\beta_{\text{Effort}}$ or $\beta_{\text{Reward}}$ (left and bottom panels). Significant clusters were mapped onto white matter tracts and characterized by MNI coordinates and their metrics were used in a predictive machine learning analysis (right panel; see D.). D. Machine learning analysis. Identified microstructural features were used as inputs for machine learning classifiers predicting low and high $\beta_{\text{Effort}}$ and $\beta_{\text{Reward}}$ . Classifier performance was assessed using 5-fold cross-validation, yielding average accuracy and area under the receiving operator curve metrics.

Figure 4.2 : Effort and reward sensitivity show substantial inter-individual variability across healthy subjects. A. Individual acceptance rate maps. (Caption continued next page)

MNI coordinates			Volume		Partial correlations
x	y	z	(mm 3 )	Tract label	R	p
5 clusters: white matter integrity and $\boldsymbol{\beta}_{\textbf{Effort}}$
FA analysis
-13.95	-10.01	58.02	94	SMA portion of the left corticospinal tract	-0.697	$<$ 0.001
-10.93	15.96	19.02	56	Left mid-ant. segment of the corpus callosum	-0.556	$<$ 0.001
-9.95	34.7	11.22	209	Left anterior cingulum bundle	0.623	$<$ .001

Table 4.1: Clusters showing significant associations between white matter integrity and effort/reward

Figure 4.4 : Microstructure in tracts connected to frontal valuation regions showing negative associations with effort sensitivity ( $\beta_{\text{Effort}}$ ). (Caption continued on next page.)

Figure 4.6 : Microstructure in a tract connected to frontal valuation regions showing a positive association with effort sensitivity ( $\beta_{\text{Effort}}$ ). (i) A significant cluster (green) exhibited a positive correlation between FA and $\beta_{\text{Effort}}$ (MNI: -9.95, 34.7, 11.22). (ii) Probabilistic tract labelling identified the anterior cingulum. To clarify which medial frontal regions were specifically connected by this cluster, we projected it onto a high-resolution atlas, revealing fibers linking the dACC with SMA and medial OFC, core valuation hubs. (iii) Partial correlation confirmed that greater FA in this cluster was associated with higher effort sensitivity (R = 0.623, $p<0.001$ ). No additional clusters were identified in major tracts, and the -MD analysis did not reveal any clusters with positive associations.

Figure 4.7 : Microstructure in tracts connected to frontal valuation regions and sensorimotor structures and showing negative associations with reward sensitivity ( $\beta_{\text{Reward}}$ ) (Figure caption continued on next page.)

Figure 4.9 : Exclusive overlap of white matter clusters associated with effort and reward sensitivity within the SMA portion of the corticospinal tract. FA analyses identified two significant clusters within the SMA portion of the left corticospinal tract. The cluster associated with $\beta_{\text{Effort}}$ (red) was centered at MNI: X = –13.95, Y = –10.01, Z = 58.02 (94 mm³), while the cluster associated with $\beta_{\text{Reward}}$ (green) was centered at MNI: X = –13.81, Y = –10.44, Z = 59.95. These clusters overlapped extensively (yellow), forming the only locus across the entire FA and -MD maps where microstructural integrity covaried with both $\beta_{\text{Effort}}$ and $\beta_{\text{Reward}}$ . This convergence underscores the specificity of this SMA corticospinal segment as a shared structural substrate for individual differences in sensitivity to both effort and reward.

Figure 4.10 : Microstructure in tracts connected to frontal valuation regions, fronto-parietal and sensorimotor structures showing positive associations with reward sensitivity. (Caption continued on next page.)

Figure 4.12 : White matter microstructure predicts individual differences in effort and reward sensitivity. (Caption continued next page.)

Figure 4.14 : Decoding effort and reward sensitivity is classifier-independent and cluster-specific. A. Classifier-independent and cluster-specific decoding of $\beta_{\text{Effort}}$ . Left panel (blue): Area under the curve (AUC) values for 12 machine learning classifiers trained on microstructural measures from the 5 clusters significantly associated with $\beta_{\text{Effort}}$ . Decoding performance was significantly above chance for nearly all classifiers (all FDR-corrected $p<0.001$ ), except for SVM (FDR-corrected $p=0.99$ ), demonstrating classifier-independent decoding. Classifier names: LR: logistic regression; SVM: support vector machine; RF: random forest; GB: gradient boosting; ET: ExtraTrees; AdaB: AdaBoost; DT: decision tree; MLP: multilayer perceptron; KNN: k-nearest neighbors; GNB: Gaussian naive Bayes; LDA: linear discriminant analysis; QDA: quadratic discriminant analysis. Middle panel (pink): AUC values for the same classifiers trained on microstructural measures from randomly sampled white matter clusters, matched in size and number to the original clusters, showing performance near chance. Right panel: Statistical comparison of AUC values from the best-performing classifier (ExtraTrees) trained on the 5 significant clusters (blue) versus 5 random clusters (pink). Performance with significant clusters was significantly higher than with random clusters (***FDR-corrected $p<0.001$ ) and above chance (###FDR-corrected $p<0.001$ ), whereas performance with random clusters did not differ from chance (n.s.). Box plots indicate median and interquartile range; whiskers denote data within $1.5\times\text{IQR}$ . B. Classifier-independent and cluster-specific decoding of $\beta_{\text{Reward}}$ . The panels follow the same organization as A. but for $\beta_{\text{Reward}}$ .

Figure 4.15 : SMA-connected clusters dominate in terms of predictive power, with distributed circuits contributing to reward sensitivity decoding. A. Cluster contributions and optimal cluster set for decoding $\beta_{\text{Effort}}$ . Left panel: Feature importance analysis, where “features” denote the individual white matter clusters used as input variables in the model (a standard machine learning term for predictors). The most predictive cluster was the left corticospinal tract, SMA portion [FA], followed by the right and left premotor portions of the corpus callosum [-MD and FA], and the left anterior cingulum [FA]. Right panel: Recursive feature elimination analysis. Classification performance (AUC) peaked when restricted to the four most informative clusters (green diamond), which are highlighted By the green dashed box in the left panel. B. Cluster contributions and optimal cluster set for decoding $\beta_{\text{Reward}}$ . The panels follow the same organization as A. but for $\beta_{\text{Reward}}$ . Here, the left corticospinal tract, SMA portion [FA], was again the most predictive feature, followed by clusters in the cortico-ponto-cerebellar tract [-MD], superior lateral fasciculus [FA], and forceps minor [FA]. Maximal classification performance (AUC) was also reached with four features, indicating that reward sensitivity is best predicted by a combination of SMA-connected, fronto-parietal, and cerebellar pathways.

Figure 5.1 : GMV from structural MRI and correlational analysis method. A. Structural MRI data were acquired from 45 healthy participants and processed using FreeSurfer for volumetric segmentation and the Brainnetome Atlas for parcellation (Caption continued next page).

Effort sensitivity
Region	MNI coordinates (x, y, z)	r	p
Right dorsal insula	39, $-7$ , 8	0.315	0.0400
Left postcentral superior parietal lobule	$-22$ , $-47$ , 65	0.397	0.00832
Left primary motor cortex	5, $-21$ , 61	0.372	0.0142
Right rostroventral inferior parietal lobule	55, $-26$ , 26	0.331	0.0303
Reward sensitivity

Table 5.1 : Brain regions showing significant correlations between GMV and apathy scores, effort, and reward sensitivity

Figure 5.3 : Higher effort sensitivity was associated with greater GMV in A. the right dorsal insula (MNI coordinates: 39, $-7$ , 8; $R=0.315$ , $p=0.040$ ), B. the left postcentral superior parietal lobule (MNI coordinates: $-22$ , $-47$ , 65; $R=0.397$ , $p=0.00832$ ), C. the left primary motor cortex (leg area) (MNI coordinates: 5, $-21$ , 61; $R=0.372$ , $p=0.0142$ ), and D. the right rostroventral inferior parietal lobule (MNI coordinates: 55, $-26$ , 26; $R=0.331$ , $p=0.0303$ ). Scatter plots on the right depict partial correlations between regional GMV and effort sensitivity after controlling for age, gender, depression, and anhedonia.

Figure 5.4 : Reward sensitivity ( $\beta_{\text{Reward}}$ ) showed positive correlations with GMVs in A. the right postcentral area of the superior parietal lobule (MNI coordinates: 23, $-43$ , 67; $R=0.428$ , $p=0.00422$ ) and B. the left ventromedial putamen (MNI coordinates: $-23$ , 7, $-4$ ; $R=0.356$ , $p=0.0193$ ). Individuals with lower regional volume in these areas exhibited reduced reward sensitivity. Scatter plots show partial correlations after controlling for age, gender, depression, and anhedonia.

Figure 5.5 : Apathy scores and GMV correlation (Caption continued on next page)

Figure 5.7 : Classifier performance compared to chance level across classification tasks. A, B. For effort sensitivity and reward sensitivity classifications, all classifiers except the support vector machine (SVM) achieved performance significantly higher than chance. C. In the apathy level classification, all models performed significantly above chance. Boxplots show the distribution of the area under the receiver operating characteristic curve (AUC) across cross-validation folds for each model. The orange line indicates the theoretical chance level (AUC = 0.5). ***p $<$ 0.001.

Figure 5.8 : Permutation test on apathy score, effort sensitivity and reward sensitivity classification performance. Caption continued on next page.

Figure 5.9 : A. Effort sensitivity classification. (i) Classifier accuracy for true-label data (blue line) did not differ significantly from the permuted-label distribution (pink; Monte Carlo $p=0.548$ ). (ii) Accuracy for true-label data was higher than chance but not significantly different from permuted-label data (true vs. permuted: $t_{998}=1.13$ , $p=0.26$ ; true vs. chance: $t_{999}=48.19$ , $p<0.001$ ). (iii) ROC curve showing mean AUC = $0.621\pm 0.035$ (95% CI). (iv) True-label AUC did not differ significantly from the permuted-label distribution (Monte Carlo $p=0.101$ ). (v) AUC was significantly higher than chance but not from permuted data (true vs. permuted: $t_{998}=1.71$ , $p=0.087$ ; true vs. chance: $t_{999}=54.62$ , $p<0.001$ ). B. Reward sensitivity classification. (i) Classifier accuracy on true-label data (blue line) was significantly higher than the null distribution from 1,000 permutations (pink); only 17 permutations matched or exceeded true-label accuracy (Monte Carlo $p=0.017$ ). (ii) Accuracy was significantly higher for true-label data (true vs. permuted: $t_{998}=44.83$ , $p<0.001$ ; true vs. chance: $t_{999}=89.72$ , $p<0.001$ ). (iii) ROC curve showing classifier performance (mean AUC = $0.705\pm 0.047$ ). (iv) True-label AUC exceeded the permuted-label distribution (Monte Carlo $p=0.044$ ). (v) AUC was significantly higher for true-label data compared to permuted and chance-level data (true vs. permuted: $t_{998}=39.62$ , $p<0.001$ ; true vs. chance: $t_{999}=96.84$ , $p<0.001$ ). Boxplots show median, interquartile range, and full range. Symbols denote statistical significance: ### $p<.001$ , *** $p<.001$ , n.s. = not significant. Caption continued on next page.

Figure 5.11 : Feature importance analysis of the best classifiers among three classification experiments. Feature importance values were derived from the best-performing classifiers predicting individual differences in apathy, effort sensitivity, and reward sensitivity from regional GMVs. A. For apathy classification, the ventrolateral fusiform gyrus and supplementary motor area showed the strongest contributions. B. For effort sensitivity classification, the dorsal insula and superior parietal lobule were the most predictive regions. C. For reward sensitivity classification, the ventromedial putamen and superior parietal lobule contributed most strongly to model performance.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

5.1.Figure : Ablation study of EEG feature contributions. (A) Resting-state EEG. Removing motor electrode features led to a significant reduction in AUC compared with the baseline model (*p $<$ 0.05), whereas removing the top-ranked features identified by feature importance analysis did not significantly affect performance. (B) Stop-signal task EEG. Removing motor electrode features did not alter performance, but removing top-ranked features significantly reduced AUC compared with both baseline and motor-electrode ablation (***p $<$ 0.001). Error bars indicate ±1 SEM.

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.535 ± 0.030	0.628 ± 0.015	0.011 ± 0.024	0.027 ± 0.055	0.013 ± 0.026
SVM	0.464 ± 0.042	0.623 ± 0.020	0.018 ± 0.027	0.033 ± 0.048	0.021 ± 0.029
RF	0.519 ± 0.033	0.588 ± 0.025	0.171 ± 0.046	0.352 ± 0.108	0.216 ± 0.055
GB	0.495 ± 0.047	0.585 ± 0.033	0.188 ± 0.064	0.353 ± 0.123	0.225 ± 0.068
ET	0.533 ± 0.035	0.602 ± 0.025	0.130 ± 0.045	0.305 ± 0.121	0.169 ± 0.056
AdaB	0.533 ± 0.043	0.593 ± 0.035	0.271 ± 0.060	0.416 ± 0.096	0.307 ± 0.060

Appendix

5.1.Table : Appendix table: ADHD classification performance with resting-state EEG data when removing top 5 most important electrodes from feature importance analysis

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.538 ± 0.032	0.628 ± 0.016	0.013 ± 0.024	0.026 ± 0.052	0.014 ± 0.025
SVM	0.471 ± 0.042	0.622 ± 0.021	0.013 ± 0.023	0.026 ± 0.046	0.015 ± 0.024
RF	0.517 ± 0.032	0.591 ± 0.025	0.163 ± 0.049	0.357 ± 0.112	0.208 ± 0.058
GB	0.505 ± 0.043	0.589 ± 0.037	0.194 ± 0.065	0.372 ± 0.130	0.234 ± 0.070
ET	0.527 ± 0.038	0.600 ± 0.025	0.117 ± 0.044	0.288 ± 0.128	0.155 ± 0.057
AdaB	0.513 ± 0.036	0.582 ± 0.035	0.254 ± 0.052	0.390 ± 0.093	0.287 ± 0.050

Appendix

5.2.Table : Appendix table: ADHD classification performance with resting-state EEG data when removing motor electrodes

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.639 ± 0.062	0.834 ± 0.015	0.977 ± 0.019	0.850 ± 0.009	0.908 ± 0.009
SVM	0.456 ± 0.095	0.834 ± 0.015	0.981 ± 0.022	0.847 ± 0.007	0.908 ± 0.010
RF	0.613 ± 0.047	0.829 ± 0.015	0.966 ± 0.013	0.852 ± 0.008	0.905 ± 0.009
GB	0.564 ± 0.058	0.837 ± 0.013	0.991 ± 0.016	0.843 ± 0.003	0.911 ± 0.008
ET	0.637 ± 0.041	0.828 ± 0.011	0.975 ± 0.012	0.845 ± 0.007	0.905 ± 0.007
AdaB	0.642 ± 0.059	0.815 ± 0.024	0.952 ± 0.024	0.848 ± 0.011	0.896 ± 0.014

Appendix

5.3.Table : Appendix table: ADHD classification performance with stop signal task EEG data when removing top 5 most important electrodes from feature importance analysis

LR	0.698 ± 0.059	0.833 ± 0.019	0.967 ± 0.025	0.856 ± 0.013	0.906 ± 0.012
SVM	0.474 ± 0.099	0.831 ± 0.014	0.974 ± 0.021	0.849 ± 0.010	0.906 ± 0.009
RF	0.635 ± 0.049	0.835 ± 0.014	0.970 ± 0.011	0.855 ± 0.009	0.908 ± 0.008
GB	0.684 ± 0.054	0.836 ± 0.016	0.980 ± 0.021	0.849 ± 0.009	0.909 ± 0.010
ET	0.645 ± 0.047	0.831 ± 0.011	0.978 ± 0.009	0.846 ± 0.007	0.907 ± 0.006
AdaB	0.768 ± 0.045	0.841 ± 0.018	0.958 ± 0.018	0.869 ± 0.012	0.910 ± 0.011
DT	0.629 ± 0.056	0.794 ± 0.039	0.876 ± 0.043	0.881 ± 0.018	0.876 ± 0.027

Appendix

5.4.Table : Appendix table: ADHD classification performance with stop signal task EEG data when removing motor electrodes

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.652 $\pm$ 0.053	0.613 $\pm$ 0.042	0.580 $\pm$ 0.063	0.644 $\pm$ 0.075	0.586 $\pm$ 0.058
SVM	0.490 $\pm$ 0.099	0.561 $\pm$ 0.055	0.484 $\pm$ 0.096	0.590 $\pm$ 0.109	0.497 $\pm$ 0.085
RF	0.705 $\pm$ 0.052	0.630 $\pm$ 0.054	0.612 $\pm$ 0.060	0.665 $\pm$ 0.080	0.613 $\pm$ 0.056
GB	0.651 $\pm$ 0.068	0.601 $\pm$ 0.060	0.603 $\pm$ 0.080	0.627 $\pm$ 0.081	0.592 $\pm$ 0.067
ET	0.719 $\pm$ 0.050	0.648 $\pm$ 0.053	0.602 $\pm$ 0.064	0.698 $\pm$ 0.084	0.619 $\pm$ 0.059
AdaB	0.644 $\pm$ 0.055	0.568 $\pm$ 0.051	0.601 $\pm$ 0.079	0.588 $\pm$ 0.072	0.568 $\pm$ 0.061

Appendix

6.1.Table: Decoding effort sensitivity is classifier-independent: further evidence from multiple machine learning metrics. Performance of 12 machine learning classifiers in predicting effort sensitivity from white matter microstructural features, including fractional anisotropy and mean diffusivity. Reported values reflect the mean ± standard deviation across 1,000 iterations of stratified 5-fold cross-validation with randomized data partitioning. Classifier abbreviations: LR, logistic regression; SVM, support vector machine; RF, random forest; GB, gradient boosting; ET, extra trees; AdaB, AdaBoost; DT, decision tree; MLP, multi-layer perceptron; KNN, k-nearest neighbors; GNB, Gaussian naïve Bayes; LDA, linear discriminant analysis; QDA, quadratic discriminant analysis. Performance metrics include area under the receiver operating characteristic curve (AUC), accuracy, recall (sensitivity), precision (positive predictive value), and F1 score (harmonic mean of precision and recall). Among classifiers, the ET model achieved the highest AUC (0.719 ± 0.050), along with a balance between precision (0.698 ± 0.084) and F1 score (0.619 ± 0.059), indicating robust discriminative performance. RF and QDA also performed well (AUCs of 0.705 ± 0.052 and 0.682 ± 0.055, respectively), though QDA exhibited reduced recall (0.430 ± 0.062), suggesting a trade-off favoring specificity. Linear models such as LR and LDA demonstrated moderate AUCs (0.652 ± 0.053 and 0.671 ± 0.048, respectively), with LDA yielding relatively high precision (0.679 ± 0.065). In contrast, SVM showed the lowest AUC (0.490 ± 0.099), indicating performance at or below chance levels. GNB and QDA showed the largest discrepancies between precision and recall, highlighting imbalances in classification. Hence, despite variations in individual metrics, most classifiers achieved above-chance performance, reinforcing the robustness of the decoding. These results collectively indicate that effort sensitivity can be reliably predicted from white matter features across a wide range of machine learning approaches.

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.837 $\pm$ 0.034	0.718 $\pm$ 0.038	0.749 $\pm$ 0.052	0.736 $\pm$ 0.047	0.722 $\pm$ 0.043
SVM	0.695 $\pm$ 0.117	0.704 $\pm$ 0.041	0.787 $\pm$ 0.064	0.707 $\pm$ 0.049	0.721 $\pm$ 0.046
RF	0.763 $\pm$ 0.042	0.678 $\pm$ 0.039	0.689 $\pm$ 0.059	0.702 $\pm$ 0.053	0.673 $\pm$ 0.047
GB	0.711 $\pm$ 0.058	0.675 $\pm$ 0.054	0.702 $\pm$ 0.083	0.691 $\pm$ 0.064	0.674 $\pm$ 0.067
ET	0.807 $\pm$ 0.037	0.693 $\pm$ 0.040	0.743 $\pm$ 0.057	0.707 $\pm$ 0.051	0.702 $\pm$ 0.045
AdaB	0.774 $\pm$ 0.048	0.710 $\pm$ 0.051	0.773 $\pm$ 0.073	0.712 $\pm$ 0.060	0.720 $\pm$ 0.058

Appendix

6.2.Table: Decoding reward sensitivity is classifier-independent: further evidence from multiple machine learning metrics. Performance of 12 machine learning classifiers in predicting reward sensitivity from white matter microstructural features, including fractional anisotropy and mean diffusivity. Reported values reflect the mean ± standard deviation across 1,000 iterations of stratified 5-fold cross-validation with randomized data partitioning. Classifier abbreviations: LR, logistic regression; SVM, support vector machine; RF, random forest; GB, gradient boosting; ET, extra trees; AdaB, AdaBoost; DT, decision tree; MLP, multi-layer perceptron; KNN, k-nearest neighbors; GNB, Gaussian naïve Bayes; LDA, linear discriminant analysis; QDA, quadratic discriminant analysis. Performance metrics include area under the receiver operating characteristic curve (AUC), accuracy, recall (sensitivity), precision (positive predictive value), and F1 score (harmonic mean of precision and recall). Across models, LR and LDA yielded the highest AUC values (0.837 ± 0.034, highlighted in the table and 0.834 ± 0.035, respectively), indicating strong overall discriminative ability. GNB and QDA also performed competitively (AUCs of 0.802 ± 0.030 and 0.792 ± 0.049, respectively), with GNB achieving the highest recall (0.815 ± 0.042), suggesting strong sensitivity to effort-related signal in the input features. In contrast, tree-based methods such as DT and GB exhibited comparatively lower performance across most metrics, highlighting limitations in their generalizability to the underlying data structure. While MLP and KNN classifiers offered moderate performance, ensemble methods like AdaB and ET demonstrated a balance between precision and recall, though their AUCs fell below those of the linear models. These results suggest that linear decision boundaries may be particularly well suited for classifying reward sensitivity from microstructural white matter features, potentially reflecting a low-dimensional organization of the relevant neural representations. Further, despite differences in classifier type and complexity, most models achieved consistent above-chance accuracy, demonstrating that reward sensitivity is robustly decodable. These results again indicate that reward sensitivity can be reliably predicted from white matter features across a wide range of machine learning approaches.

6.1.Figure : A. Candidate models. Candidate models of subjective value computation were tested to quantify individual differences in effort and reward sensitivity. Models differed in cost function (linear, quadratic, or exponential) and in whether they included a time-dependent modulation term ( $\beta_{\text{Time}}$ ), following procedures established in prior work (e.g., Le Heron et al. [2018a] ). Value was modelled as a function of effort level ( $EFFORT_{\text{LEVEL}}$ ), reward level ( $REWARD_{\text{LEVEL}}$ ), and trial number ( $t$ ), with free parameters capturing individual sensitivities ( $\beta_{\text{Effort}}$ , $\beta_{\text{Reward}}$ ) and baseline bias ( $\beta_{0}$ ). B. Model fit comparison based on Bayesian information criterion (BIC). Model 4, which combined a quadratic cost function with a time-modulated cost term, showed the best fit (lowest BIC), consistent with prior findings ( Le Heron et al. [2018a] , Pessiglione et al. [2018] , Prévost et al. [2010] . C. Best-fitting model vs. observed behavior. Model-derived acceptance rate (black dashed lines) closely captured observed behavioral data (blue lines; $N=45$ ), across varying effort levels and reward magnitudes.

6.2.Figure : Superficial cluster showing a positive association between fractional anisotropy (FA) and $\beta_{\text{Effort}}$ . i. A significant cluster was identified outside canonical white matter tracts, located in the left parietal cortex (MNI coordinates: X = $-25.66$ , Y = $-45.76$ , Z = $38.05$ ; volume: 62 mm 3 ). ii. Partial correlation analysis confirmed this association, revealing a robust positive correlation between FA values extracted from the cluster and $\beta_{\text{Effort}}$ ( $R=.583$ , $p<.001$ ). These findings suggest that, in addition to the major tracts identified using a probabilistic white matter atlas, more superficial white matter regions may also contribute to individual variability in effort sensitivity. This further highlights the added value of whole-brain, data-driven approaches for uncovering relevant structural substrates beyond predefined tracts-of-interest.

6.3.Figure : Three clusters with positive $\beta_{\text{Reward}}$ associations in superficial white matter of the left frontal and parietal cortex outside major tracts. Three additional clusters exhibiting significant positive associations between white matter integrity and reward sensitivity ( $\beta_{\text{Reward}}$ ) were identified in superficial regions of the left frontal and parietal cortex, outside canonical major white matter tracts. A. i. Cluster (green) located at MNI coordinates: X = $-16.44$ , Y = $51.23$ , Z = $23.61$ ; volume: 88 mm 3 . ii. Partial correlation between fractional anisotropy (FA) in this cluster and $\beta_{\text{Reward}}$ : $R=.574$ , $p<.001$ . B. i. Cluster (green) located at MNI coordinates: X = $-17.64$ , Y = $44.13$ , Z = $18.88$ ; volume: 64 mm 3 . ii. Partial correlation with $\beta_{\text{Reward}}$ : $R=.546$ , $p<.001$ . C. i. Cluster (green) located at MNI coordinates: X = $-42.32$ , Y = $-60.22$ , Z = $15.16$ ; volume: 64 mm 3 ). ii. Partial correlation with $\beta_{\text{Reward}}$ : $R=.629$ , $p<.001$ .

6.4.Figure : The SMA cluster associated with both $\beta_{\text{Effort}}$ and $\beta_{\text{Reward}}$ is specific and unrelated to the two other model parameters. (Figure caption continued on next page.)

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.815 ± 0.037	0.722 ± 0.041	0.735 ± 0.053	0.746 ± 0.053	0.721 ± 0.045
SVM	0.786 ± 0.052	0.704 ± 0.047	0.641 ± 0.066	0.769 ± 0.072	0.674 ± 0.058
RF	0.739 ± 0.044	0.685 ± 0.039	0.681 ± 0.049	0.717 ± 0.054	0.678 ± 0.043
GB	0.656 ± 0.058	0.616 ± 0.051	0.614 ± 0.076	0.639 ± 0.070	0.605 ± 0.062
ET	0.755 ± 0.043	0.691 ± 0.042	0.664 ± 0.053	0.736 ± 0.061	0.676 ± 0.046
AdaB	0.714 ± 0.047	0.647 ± 0.043	0.595 ± 0.063	0.697 ± 0.072	0.614 ± 0.055

Appendix

7.1.Table: Appendix table: Apathy scores ML classification performance with grey matter volumes

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.562 ± 0.055	0.631 ± 0.036	0.100 ± 0.074	0.158 ± 0.132	0.114 ± 0.085
SVM	0.470 ± 0.077	0.617 ± 0.038	0.053 ± 0.055	0.084 ± 0.096	0.060 ± 0.062
RF	0.568 ± 0.058	0.638 ± 0.045	0.317 ± 0.079	0.426 ± 0.129	0.340 ± 0.083
GB	0.541 ± 0.068	0.608 ± 0.051	0.259 ± 0.098	0.333 ± 0.141	0.271 ± 0.100
ET	0.553 ± 0.060	0.616 ± 0.041	0.130 ± 0.073	0.223 ± 0.143	0.153 ± 0.086
AdaB	0.564 ± 0.057	0.588 ± 0.049	0.240 ± 0.099	0.271 ± 0.125	0.236 ± 0.095

Appendix

7.2.Table: Appendix table: Effort sensitivity ML classification performance with grey matter volumes

Model	AUC	Accuracy	Recall	Precision	F1
LR	0.649 ± 0.050	0.605 ± 0.038	0.405 ± 0.080	0.563 ± 0.126	0.445 ± 0.085
SVM	0.487 ± 0.101	0.591 ± 0.046	0.298 ± 0.078	0.556 ± 0.153	0.359 ± 0.084
RF	0.646 ± 0.052	0.591 ± 0.046	0.481 ± 0.071	0.556 ± 0.093	0.490 ± 0.066
GB	0.670 ± 0.060	0.627 ± 0.050	0.509 ± 0.081	0.624 ± 0.105	0.526 ± 0.069
ET	0.630 ± 0.048	0.580 ± 0.044	0.400 ± 0.066	0.544 ± 0.106	0.433 ± 0.064
AdaB	0.685 ± 0.051	0.673 ± 0.044	0.471 ± 0.081	0.727 ± 0.124	0.529 ± 0.072

Appendix

7.3.Table: Appendix table: Reward sensitivity ML classification performance with grey matter volumes

Explore similar work

May 18, 2026

Subject-Specific Analysis of Self-Initiated Attention Shifts from EEG with Controlled Internal and External Attention Conditions

Self-initiated attention shifts play a critical role in voluntary behavior but are difficult to study due to the absence of explicit temporal markers. While previous studies have examined their neural correlates, it remains unclear how multi-dimensional electroencephalography (EEG) features contribute to their characterization within an interpretable computational framework. In this study, we build on an experimental paradigm developed in our previous work, which enables controlled comparison between task-constrained self-initiated shifts and externally instructed shifts under identical visual stimulation. Within this setting, we investigate whether preparatory EEG activity can distinguish these two types of attention shifts. We adopt a machine learning-based approach and conduct two complementary analyses: (1) a performance-oriented assessment of frequency-specific topographic patterns, and (2) a model-based feature attribution analysis using SHapley Additive exPlanations (SHAP). These analyses provide a structured view of how spectral features across regions of interest contribute to model behavior. Our results demonstrate reliable within-subject classification performance, indicating that preparatory EEG activity contains subject-specific discriminative information within this paradigm. The analysis shows that higher-frequency bands and frontal regions contribute strongly to model decisions, although such contributions should be interpreted cautiously due to the potential influence of non-neural artifacts in high-frequency EEG signals. Overall, this work highlights the value of interpretable machine learning for analyzing subject-specific EEG signal patterns in a controlled experimental setting, with potential applications in personalized and asynchronous brain-machine interface systems.

Yuwen Zeng, Dengzhe Hou, Zhang Zhang +4

Advanced Institute of Convergence Knowledge Informatics, Tohoku University, Japan · Graduate School of Information Sciences, Tohoku University, Japan · Center for Data-Driven Science and Artificial Intelligence, Tohoku University, Japan +1

Electroencephalography EEG Decoding

Ask PDF

Mar 3, 2026

Stringology-Based Motif Discovery from EEG Signals: an ADHD Case Study

We propose a novel computational framework for analyzing electroencephalography (EEG) time series using methods from stringology, the study of efficient algorithms for string processing, to systematically identify and characterize recurrent temporal patterns in neural signals. The primary aim is to introduce quantitative measures to understand neural signal dynamics, with the present findings serving as a proof-of-concept. The framework adapts order-preserving matching (OPM) and Cartesian tree matching (CTM) to detect temporal motifs that preserve relative ordering and hierarchical structure while remaining invariant to amplitude scaling. This approach provides a temporally precise representation of EEG dynamics that complements traditional spectral and global complexity analyses. To evaluate its utility, we applied the framework to multichannel EEG recordings from individuals with attention-deficit/hyperactivity disorder (ADHD) and matched controls using a publicly available dataset. Highly recurrent, group-specific motifs were extracted and quantified using both OPM and CTM. The ADHD group exhibited significantly higher motif frequencies, suggesting increased repetitiveness in neural activity. OPM analysis revealed shorter motif lengths and greater gradient instability in ADHD, reflected in larger mean and maximal inter-sample amplitude changes. CTM analysis further demonstrated reduced hierarchical complexity in ADHD, characterized by shallower tree structures and fewer hierarchical levels despite comparable motif lengths. These findings suggest that ADHD-related EEG alterations involve systematic differences in the structure, stability, and hierarchical organization of recurrent temporal patterns. The proposed stringology-based motif framework provides a complementary computational tool with potential applications for objective biomarker development in neurodevelopmental disorders.

Anat Dahan, Samah Ghazawi

Braude College of Engineering, Karmiel, Israel

Electroencephalography Mental Health

Ask PDF

Mar 27, 2026

D-GATNet: Interpretable Temporal Graph Attention Learning for ADHD Identification Using Dynamic Functional Connectivity

Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder whose neuroimaging-based diagnosis remains challenging due to complex time-varying disruptions in brain connectivity. Functional MRI (fMRI) provides a powerful non-invasive modality for identifying functional alterations. Existing deep learning (DL) studies employ diverse neuroimaging features; however, static functional connectivity remains widely used, whereas dynamic connectivity modeling is comparatively underexplored. Moreover, many DL models lack interpretability. In this work, we propose D-GATNet, an interpretable temporal graph-based framework for automated ADHD classification using dynamic functional connectivity (dFC). Sliding-window Pearson correlation constructs sequences of functional brain graphs with regions of interest as nodes and connectivity strengths as edges. Spatial dependencies are learned via a multi-layer Graph Attention Network, while temporal dynamics are modeled using 1D convolution followed by temporal attention. Interpretability is achieved through graph attention weights revealing dominant ROI interactions, ROI importance scores identifying influential regions, and temporal attention emphasizing informative connectivity segments. Experiments on the Peking University site of the ADHD-200 dataset using stratified 10-fold cross-validation with a 5-seed ensemble achieved 85.18% +_5.64 balanced accuracy and 0.881 AUC, outperforming state-of-the-art methods. Attention analysis reveals cerebellar and default mode network disruptions, indicating potential neuroimaging biomarkers.

Qurat Ul Ain, Alptekin Temizel, Soyiba Jawed

Department of Computer and Software Engineering, College of Electrical and Mechanical Engineering, National University of Sciences and Technology, Islamabad 44000, Pakistan · Graduate School of Informatics, Middle East Technical University, 06800, Ankara, Turkey

Classification Neuroimaging Graph Attention Networks

Ask PDF