A total of 35 children with ADHD and 28 HCs (7–14 years old) were recruited in the child dataset. The diagnosis was made by a senior psychiatrist based on the Schedule for Affective Disorders and Schizophrenia for School-Age Children-Present and Lifetime version (K-SADS-PL)15, a clinical and semi-structured interview based on Diagnostic and Statistical Manual of Mental Disorders-Fourth Edition (DSM-IV)16.
A total of 112 adults with ADHD and 77 age- and sex-matched HCs (18–40 years old) were recruited in the adult dataset. The Conner’s Adult ADHD Diagnostic Interview based on DSM-IV was completed for the diagnosis of adults with ADHD. All adult participants also underwent the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I)17 by a senior psychiatrist for potential comorbidity.
All participants met the following criteria: (1) right-handed, (2) no history of head trauma with a loss of consciousness, (3) no history of neurological disorders or other severe disease, (4) no current diagnosis of major depressive disorder (MDD), schizophrenia, clinically significant panic disorder, bipolar disorder, pervasive developmental disorders, or mental retardation, (5) no excessive head movements (>3.0 mm of translation or degrees of rotation in any direction), and (6) a full-scale intelligence quotient (IQ) above 80. Furthermore, participants with any history of psychiatric disorders were also excluded as HCs.
ADHD patients were recruited from outpatient clinics of Peking University Sixth Hospital and HCs were recruited by advertisement. Adult participants were scanned in Peking University Sixth Hospital (73 ADHD, 43 HCs) and Beijing Normal University (39 ADHD, 34 HCs). All child participants were scanned in Beijing Normal University. Verbal IQ, performance IQ, and full-scale IQ were measured by the Wechsler Child/Adult Intelligence Scale, Third Edition.
The severity of inattentive symptoms, hyperactive/impulsive symptoms and total ADHD symptoms of all subjects were evaluated by the ADHD Rating Scale-IV (ADHD RS-IV)18, rating one-four (“never” is rated as 1, “occasionally” is 2; “often” is 3; “always” is 4). This scale contains nine inattentive and nine hyperactive/impulsive symptoms of ADHD described in the DSM-IV. The higher the scores were, the more serious the ADHD symptoms were.
Besides, impulsivity–hyperactivity and hyperactivity index in the Conners’ Parent Rating Scale (CPRS)19 were used to assess the hyperactive/impulsive symptoms in child participants with ADHD. The CPRS is a widely used instrument for screening and evaluating ADHD-related symptoms as well as other behavioral problems frequently associated with ADHD in children. It contains 48 items and can be divided into six factors: conduct problems, learning problems, psychosomatic problems, impulsivity–hyperactivity, anxiety, and hyperactivity index. The parents rate each item using a 4-point Likert-type scale (“never/seldom” is rated as 0, “sometimes” is 1; “quite often” is 2, and “very often” is 3). The higher the score, the more severe the corresponding problem is.
This study was approved by the Research Ethics Review Board of Peking University Sixth Hospital and Beijing Normal University. All subjects provided written informed consent and were fully informed about the research.
Resting-state functional connectivity analysis
Specific parameters for scanning were shown in the supplement. Two datasets were both preprocessed using the Data Processing Assistant for Resting-State fMRI20 (DPARSFA, http://rfmri.org/DPARSF). The first 10 volumes were discarded to allow for magnetization equilibrium. Subsequent data preprocessing included slice timing correction, head motion correction, spatial normalization to the MNI template, resampling to 3 × 3 × 3 mm3, smoothing using a 4 mm Gaussian kernel, temporal band-pass filtering (0.01 Hz to 0.1 Hz), nuisance signal regression (including six head motion parameters, white matter, cerebrospinal fluid, and global signals), and head motion scrubbing (the mean frame-wise displacement, as described by Jenkinson et al.21). The registered fMRI volumes with the MNI template were divided into 273 regions according to the Brainnetome Atlas22 incorporating 210 cortical, 36 subcortical, and 27 cerebellar regions.
Regional mean time series were obtained for each by averaging the fMRI time series over all voxels in each of the 273 regions. Pearson correlation coefficients between pairs of node time courses were calculated and normalized to z score using Fisher transformation, resulting in a 273 × 273 symmetric connectivity matrix for each subject. Removing 273 diagonal elements, we extracted the upper triangle elements of the functional connectivity matrix as prediction features, i.e., the feature space for prediction was spanned by the (273 × 272)/2 = 37,128 dimensional feature vectors. In this study, FC features were described as inter-network FCs and intra-network FCs. According to the study of Yeo and his colleagues23, brain regions can be grouped into seven functional networks for visualization: visual network (VN), somatomotor network (SMN), dorsal attention network (dATN), ventral attention network (vATN), limbic network (LN), frontoparietal network (FPN), and default mode network (DMN).
Ensemble feature selection algorithm
To investigate diagnostic features between ADHD and HCs, FS_RIEL was proposed to extract most-discriminative features from high-dimensional FCs and improve the result interpretability by estimating the relative importance of features, which refers to the degree of features (i.e., FC node) contribute to classification. Five different algorithms including extreme gradient boosting24, randomized decision trees (a.k.a. ExtraTrees25), Random Forest26, AdaBoost27, and Gradient Boosting28 were employed to select features with the top 2% relative importance from different models respectively, which were assembled into a feature pool without any repetition. Then, all pooled features were fed into a linear support vector machine with a forward-backward searching strategy (SVM-FoBa)29, obtaining a more refined feature subspace. After that, the label of each subject was calculated by majority voting of the 5 base classifiers used in training. We adopted nested 3-fold cross-validation on the whole training and validation set. Please see more details in Supplementary Figs. S1 and S2 on method flowchart. To verify the validity of features we selected as well as the classification performance, we also compared the proposed model with four traditional methods including Lasso30, ElasticNet31, Fisher-Score32, and Trace_Ratio33 (Kolmogorov–Smimov test) on the classification accuracy, sensitivity, specificity, and feature dimension.
Within-cohort classification and cross-cohort prediction
We first performed the within-cohort classification as mentioned above to extract FCs that can discriminate ADHD from age-matched controls within child dataset and adult dataset, respectively. Then the identified ADHD-discriminative FCs were compared between child and adult cohorts, resulting in the shared and age-specific FCs between ADHDchild and ADHDadult. Next, the correlation between ADHD symptoms and the selected shared and distinct FC patterns were further calculated, deriving out some interesting observations.
Moreover, to investigate the stability/continuity in diagnosing of ADHD from childhood to adulthood, a cross-cohort prediction using the identified FCs in discovery cohort were performed. Namely, classifying ADHDchild from HCchild using features extracted from the ADHDadult cohort classification, and vice versa. Since the participants in child dataset were all boys, hence only male participants (74 male ADHD adult patients and 43 age-matched HCs) were selected for the cross-cohort prediction. The flowchart of the whole study design was shown in Fig. 1.
In order to make full use of the data set, a k-fold cross-validation strategy was used to estimate the performance of discriminating ADHD from HCs (accuracy, sensitivity, and specificity). The steps of k-fold cross-validation are as follows: the data set is divided into k parts and taken turns using k-1 parts as a training set and the remaining one part as a test set. After looping k times, all subjects were guaranteed to be used in test set independently, and finally, the averaged classifying accuracy of k times is regarded as the whole classification accuracy34. Specific, every looping was repeated 10 times in this study to ensure the stability of the results. K-fold cross validation is a resample procedure usually used to evaluate the classification performance of the model on a limited dataset, which can reduce the over fitting to a certain extent. After mean value across all k trials is computed, this scheme matters less which part of subjects are used as training set and which are used as testing set. Further, it can also improve the stability of the results and obtain as much effective information as possible from the limited data. Here, 5-fold cross-validation was used in child dataset (Fig. 2a) and 10-fold cross-validation was used to in adult dataset (Fig. 2b), as the number of children was relatively small.