Functional MRI-Based Discrimination of Adhd: Insights from Working Memory and Reward Processing

and continues into adolescence and adulthood. This disorder is characterized by symptoms such as inattention, impulsivity, excessive motor activity, and restlessness. The objective of this study was to investigate distinct patterns of brain activity in children with ADHD compared to typically developing children during tasks involving working memory and reward processing. We also assessed the performance of the Support Vector Machine (SVM) in diagnosing ADHD using resting-state functional Magnetic Resonance Imaging (fMRI) data. Our dataset included fMRI data from 79 individuals, with 35 diagnosed with ADHD and 44 without the disorder, collected during working memory and reward tasks. The results revealed that the 1-back task elicited increased activation in the left hemisphere, particularly in the left middle frontal gyrus, while the 2-back task resulted in higher activation in the right hemisphere, specifically in the right superior parietal lobule. Children with ADHD exhibited decreased reward anticipation activity within the ventral striatum. Using SVM with feature selection, we identified the 500 most significant features and achieved an accuracy of 67.57% in diagnosing ADHD. This research enhances our understanding of the neural correlates of ADHD and offers insights into the potential utility of fMRI and machine learning in its diagnosis.

Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterized by persistent patterns of inattention, hyperactivity, and impulsivity, which often begin in childhood and can persist into adulthood (American Psychiatric Association, 2013). ADHD is a highly prevalent condition, affecting approximately 5-10% of children and adolescents worldwide [1]. Individuals with ADHD often experience difficulties in various aspects of their daily lives, including academic and occupational performance, interpersonal relationships, and overall quality of life. In recent years, the application of fMRI to the study of ADHD has advanced our understanding of its neurobiological basis. Research in ADHD has revealed differences in brain structure and function between individuals with ADHD and typically developing individuals, highlighting the importance of specific brain regions and networks involved in attention, impulse control, and executive functions. By studying these neural signatures, researchers aim to uncover potential biomarkers for the disorder, which may enhance diagnostic accuracy and guide the development of targeted interventions. Functional magnetic resonance imaging (fMRI) has been used in several studies to investigate the neural mechanisms of attention-deficit/hyperactivity disorder (ADHD) using machine learning techniques [2]. These studies have explored various resting state-fMRI features, including regional homogeneity (ReHo), fractional amplitude of low-frequency fluctuation (fALFF), functional connectivity, and voxel- and ROI-level functional networks [2]. Significantly, machine learning algorithms have the potential to identify functional patterns, such as the combined impact of many brain regions in distinguishing individuals with ADHD from control subjects. These patterns may go unnoticed when employing conventional methodologies. Eloyan et al. [3]. created a majority voting classification algorithm from four algorithms: random forest on motor cortex connectivity, SVM on major clusters, gradient boosting on decomposed functional connectivity, and gradient boosting on functional connectivity and motion parameters. The final model had 94% specificity and 21% sensitivity, and motor network connection was most essential in ADHD classification. Several SVM classification models show that the frontal, parietal, and cerebellar are more discriminative between ADHD and controls and ADHD inattentive and ADHD mixed subtypes. applied Pearson’s correlation (PC)-derived topographical information-based high-order functional connectivity (tHOFC) and dynamics-based high-order functional connectivity (dHOFC), the sparse representation (SR)-derived FCs including the group SR (GSR), the strength and similarity guided GSR (SSGSR), and sparse low-rank (SLR) to diagnose ADHD based on the Adolescent Brain and Cognitive Development (ABCD) dataset [4]. These measures are combined with multiple kernel learning (MKL) model for ADHD classification. Results showed that dHOFC achieved the best classification AUC of 0.7315, whose superior ability to identify ADHD patients. For ROI-based dHOFC, SSGSR, and SLR, the number of increased FCs in ADHD was much more than that of the decreased ones. These abnormal FCs of brain regions were within and between cerebellum network (CN), limbic network (LN), ventral attention network (VAN), frontoparietal network (FPN), and default mode network (DMN). The connections from DMN to LN, CN, FPN, and VAN are much stronger in ADHD patients than in healthy controls, and similar tendencies were reported in the previous studies. These findings imply that DMN and LN might act as two core nodes interfering connections with the regions from other brain networks in ADHD. The insula can also be considered as part of the salience network, which plays a crucial role in attention. However, few studies have investigated working memory, monetary reward, and feedback processing in typically developing children and children diagnosed with ADHD. The purpose of this study is to evaluate the feasibility of diagnosing ADHD by analyzing the fMRI data while children with and without ADHD performing a working memory task with and without rewards [5]. Feature selection library from Scikit-learn was used to reduce dimension by selecting top features and support vector machine (SVM) is used for distinguishing normal and ADHD children from resting fMRI.

The dataset includes functional magnetic resonance imaging (fMRI) data from 79 participants (35 with ADHD and 44 without ADHD) during a working memory task and a reward task [5]. The working memory task involves maintaining and manipulating information in memory, while the reward task involves anticipating and receiving rewards. The raw data include MPRAGE structural images and fMRI while performing the 8 n-back tasks, which varied in domain (spatial and cognitive/verbal), reward (small and large), and feedback (immediate and delay). During the MRI session, participants completed 8 n-back working memory tasks in the scanner, where they were presented with a series of letters and had to judge whether the letter was the same as n letters back or at the same position in letters back. The reward amount and feedback time varied for each task. Two sessions were conducted with an interval of almost 2 years. Neuroimaging and standardized testing data were collected from 79 children, aged 8.6–12.0, of which 35 had a formal diagnosis of ADHD at session T1. 48 children returned two years later to complete standardized testing at session T2. Of the participants diagnosed with ADHD, the highest percentage was among Black individuals at 40.4%, followed by Mixed race at 30%, and then White race at 28%. Other racial groups included ’American Indian’, ’Asian’, ’Native Hawaiian’, and ’Other’. In terms of medication use, 11% of participants with ADHD knew they had the condition but were not taking any medications. The reasons for this are not provided in the dataset.

The proposed model construction flow is show in Figure 1. First, we preprocess the imaging data. Second, first level and second level statistical analysis were performed on task-fMRI. Third, features were selected for rest-fMRI to reduce dimensionality, and the final step is to train the classifier.

Figure1: Flow chart of the proposed model for ADHD diagnosis

Preprocessing

Preprocessing of functional magnetic resonance imaging (fMRI) data refers to a series of steps performed to prepare the raw data acquired from the scanner for analysis. The main objective of preprocessing is to correct for various artifacts and sources of noise that can affect the quality of the data, such as head movement, physiological noise, and image distortion. The steps of preprocessing were performed on “FSL software [6]. These steps include skull stripping [7,8] motion correction (See Figure 2), slice timing correction [9], Coregistration, spatial smoothing [10], and Normalization [11]. Figure 2 shows the motion plots for significant spikes greater than half the voxel resolution and prolonged drifts larger than the size of a single voxel. If the motion is more than half a voxel relative or over a voxel absolute, it might be necessary to utilize advanced correction methods such as scrubbing or excluding the run from the analysis. In our case the motion is not more than half over a voxel absolute so we will proceed with the same configurations.

Figure 2: Motion correction graph after performing all pre-processing steps. It denotes the movement of the time-series for a particular run, with the volume index plotted on the x-axis and the motion amount (in millimeters) displayed on the y-axis.

Statistical analysis

A commonly used statistical model in fMRI is the general linear model (GLM), which models the relationship between the observed blood-oxygen- level dependent (BOLD) signal and the experimental conditions or tasks. We have used FSL (FMRIB Software Library) to perform statistical analyses. We used a general linear model (GLM) to model the effect of task conditions (working memory with and without reward) and group (children with and without ADHD) on brain activation. We also performed group-level analyses using a mixed-effects model and corrected for multiple comparisons using the family-wise error (FWE) method. First-level analysis in fMRI refers to the initial stage of statistical analysis of individual fMRI datasets, usually performed on a voxel-by-voxel basis. The purpose of first-level analysis is to identify which parts of the brain are significantly active during a given task or condition, and to compare activity between different conditions. The results of first-level analysis provide an initial assessment of which brain regions are activated during specific tasks or conditions, and they can inform the selection of regions of interest (ROIs) for further analysis. The time series in fMRI refers to the collection of brain images taken over a period of time as the subject performs a task or experiences stimuli. In fMRI, the changes in blood oxygen levels in response to neural activity are measured and transformed into images to represent brain activity. These images are collected over time, usually in the form of several volumes (image slices) per second, creating a time series of images. The time series data is then analyzed to understand changes in brain activity over time, which can provide insight into the functional organization of the brain and how it responds to various tasks and stimuli.

The second-level analysis involves examining the group-level differences in brain activity between different experimental conditions or groups of participants. In second-level analysis, statistical models are used to compare the brain activity between groups or conditions while accounting for individual variability. The most commonly used statistical models in second-level analysis are the general linear model (GLM) and the mixed-effects model. The third level analysis involves combining results from multiple subjects obtained from the second level analysis, to identify common areas of activation across the group as a whole. The purpose of the third level analysis is to investigate differences between groups or conditions at a population level, by pooling data from multiple second-level analyses. The third level analysis can involve comparing activation maps between different groups (e.g., patients vs controls) or conditions (e.g. task vs rest), and typically involves the use of statistical parametric mapping (SPM) software to perform a group level t-test or ANOVA.

The glm used in fmri can be expressed as follows:

Y(t)=X(t)β+ε(t)  - (1)

Y(t) represents the observed fMRI data at time point X(t) represents the design matrix at time point t. β represents the parameter estimates (regression coefficients) for each condition. These coefficients indicate the contribution of each experimental condition to the observed fMRI signal.

ε(t) represents the error term at time point t, which captures the noise in the fMRI data. The Figure 3 shows a design matrix for the parameters that we have set. The two columns on the right represent the ideal time- series of two different regressors in the experiment. The order of these regressors corresponds to the order in which the timing files were entered. The first column represents the ideal time-series of the1 back condition, while the second column represents the ideal time-series of the 2-back condition. The graphical representation in the text refers to a visual representation of a predicted time-series of a voxel in an fMRI experiment. It consists of a high-pass filter, which removes low frequency signals, and two columns that depict the ideal time-series for each of the two regressors in the experiment, in this case the 1 back and 2 back conditions. The red line represents the predicted time-series of the voxel based on the assumption that it is responsive to the corresponding regressor, while the white bars represent the Hemodynamic Response Function (HRF) that is convolved with the onset of each trial for that condition.

Feature Extraction

Masking is a technique used to extract specific regions of interest (ROIs) from a three-dimensional (3D) or four-dimensional (4D) dataset, such as voxel-based brain images obtained from techniques like fMRI. A mask is essentially a binary image where the voxels corresponding to the ROIs are set to 1 (indicating inclusion), and all other voxels are set to 0 (indicating exclusion). It has been implemented using python. An image of binary mask is shown below for an arbitrary subject.

Feature Selection & Classification

Total 500 features were selected using Sklearn’s feature selection (Figure 4). SVM is the data classifier which is incorporated with linear and non-linear model and this is significant with classification and prediction [12,13]. In our proposed system used linear kernel model, this model classifies individual weight of data features. Each node’s significance in linear kernel SVM, each feature is opposed to each other significance. We randomly selected 60% of data to trains the classifier and the other 40% of data are used as testing data.

Figure3: DesignMatrix

Figure4: Top 500 features/voxels as seen from three differentaxes of brain image

Activation maps were created for each task condition (1-back, 2-back, and 1-back vs 2-back) for both ADHD and control groups. These maps show the brain regions that are significantly activated or deactivated during the task compared to the baseline condition (Figure 5). Time series plots are generated for specific regions of interest (ROIs) identified from the activation maps (Figure 6). These plots show the average signal intensity of the ROI over time and can reveal how the neural activity changes throughout the task. The activation maps showed significant activation in regions associated with working memory, such as the prefrontal cortex, parietal cortex, and cingulate gyrus, for both the 1-back and 2-back tasks. The 1-back task showed greater activation in the left hemisphere, particularly in the left middle frontal gyrus, while the 2-back task showed greater activation in the right hemisphere, particularly in the right superior parietal lobule.

The contrast between 2-back and 1-back tasks showed significant activation in several regions, including the bilateral prefrontal cortex, bilateral parietal cortex, and bilateral cingulate gyrus. The ROI analysis revealed significant activation in regions associated with working memory, such as the dorsolateral prefrontal cortex and parietal cortex, for both the 1-back and 2-back tasks. The time series plots showed significant fluctuations in BOLD signal intensity in these regions during the 1-back and 2-back tasks, reflecting changes in neural activity during the working memory tasks. In the second level analysis, we found significant differences in activation between the 1-back and 2-back tasks. The 2-back task showed greater activation in the dorsolateral prefrontal cortex and the inferior parietal lobule, while the 1-back task showed greater activation in the anterior cingulate cortex. In the third level analysis, we observed that individuals with higher working memory capacity showed greater activation in the dorsolateral prefrontal cortex during the 2-back task, indicating that this region plays an important role in working memory. In the ROI analysis, we found significant activation in the dorsolateral prefrontal cortex and the inferior parietal lobule during both the 1-back and 2-back tasks. We used an ROI analysis to investigate the activation patterns in the ventral striatum, a region of the brain involved in reward processing, during the reward and no-reward conditions of the task. The ventral striatum was chosen as an ROI based on prior research indicating its involvement in reward processing [14]. Figure 7 showed lower reward anticipation activation in the ventral striatum for ADHD children. SVM classification on resting fMRI showed accuracy of about 67.57%, with a recall and precision value of about 92.92% and 41.04% respectively.

Figure5: Activation map in working memory. A. for 1-back. B. for 2-back. C. for 1-back–2-back.

Figure6: Time Series Plot

Figure7: Locationof the Ventral Striatum ROI

Present research paper delves into the realm of attention-deficit/hyperactivity disorder (ADHD) through the lens of functional magnetic resonance imaging (fMRI) and machine learning techniques. ADHD, characterized by persistent inattention, hyperactivity, and impulsivity, has substantial and enduring effects on individuals' lives. The research underscores the pivotal role of fMRI in uncovering disparities in brain structure and function between those with ADHD and their neurotypical counterparts, thereby shedding light on the neural signatures intertwined with ADHD. Significant alterations in activation patterns emerged when comparing the 1-back and 2-back tasks in the second-level analysis. Specifically, the 2-back task exhibited heightened activation in the dorsolateral prefrontal cortex and the inferior parietal lobule, whereas the anterior cingulate cortex displayed increased activation during the 1-back task. In the subsequent third- level analysis, it became evident that individuals with superior working memory capacity displayed greater activation in the dorsolateral prefrontal cortex during the 2-back task, underscoring its role in working memory processes. Furthermore, the analysis of regions of interest (ROI) during both the 1-back and 2- back tasks revealed notable activation in the dorsolateral prefrontal cortex and the inferior parietal lobule. To delve into the dynamics of reward processing, we conducted an ROI analysis focused on the ventral striatum, a brain region associated with such processes, in the context of reward and no-reward conditions throughout the task.

The study places a significant emphasis on the application of machine learning algorithms to extract functional patterns within resting-state fMRI data, potentially enabling the discernment of individuals with ADHD from control subjects. The outcomes suggest that feature selection and support vector machine (SVM) classification can effectively differentiate children with ADHD from their non-ADHD counterparts, relying on resting fMRI data. These findings provide valuable insights into the functional organization of the brain and its dynamic response to a variety of tasks and stimuli. However, reported accuracy with SVM classifier was 67.57%. Previous studies reported similar accuracies. For example, Wolfers et al. applied a Gaussian process classifier in differentiating subjects with ADHD, their unaffected siblings, and controls based on the fMRI data during stop-signal task [15]. The model was able to differentiate ADHD patients from their siblings with an AUC of 0.65 and from control participants with an AUC of 0.64. The best accuracies achieved by the ADHD-200 competition teams were 62.52% using personal characteristic data and 60.51% using fMRI data [16]. However, we have observed issue of overfitting that limit the scope for generalization. Previous study reported notable decrease in classification when a trained model is applied to unfamiliar subjects [17]. This underscores the critical necessity of tackling the issues around generalisation in the deployment of machine learning.

In essence, this research underscores the invaluable contribution of fMRI and machine learning in advancing our comprehension of the neurobiological basis of ADHD and its potential for diagnosis. By integrating advanced neuroimaging techniques and sophisticated data analysis methods, the study offers a promising path toward enhancing the accuracy of ADHD diagnosis and guiding the development of targeted interventions. There are two limitations inherent in this study. One limitation of our study is that we exclusively utilised traditional machine learning algorithms. In the future, there will be an introduction of deep learning models, such as deep models for image preprocessing and classification, which possess the ability to automatically extract feature representations that are both high-level and compact. Another limitation of our study is that we conducted binary classification of patients and controls, without considering the different subtypes of ADHD.