This study has been registered on the International Standard Randomized Controlled Trial Number Registry (ISRCTN94097629)27.
We recruited 45 adolescents (10–18 years of age, mean age 13.9 ± 1.8 years, 30 males) from a stable childcare center in New Delhi, India to participate in this research; the Udayan Care stable care center provides shelter, nurturing, and education access to children who have suffered early life adversity. Ethical approvals for this international study were obtained from the institutional review boards at the University of California San Francisco, the All India Institute for Medical Sciences (AIIMS) New Delhi, the ethical considerations for research committee at Udayan Care, as well as the Indian Health Ministry’s Screening Committee, which is the ethical body of the Indian Council of Medical Research that oversees all international collaborative research in India. Participants’ caregivers provided signed informed consent for the research, and all enrolled adolescents provided verbal assent.
All study participants completed self-reports on the childhood trauma questionnaire (CTQ35) that scores trauma in the domains of emotional and physical neglect and/or emotional, physical, and sexual abuse. We calculated individual CTQ scores as the mean scores across all five abuse/neglect subdomains that are measured on this scale (score across 45 participants, mean 2.08 ± 0.58, range 1.21–3.89). Individual subdomain scores are typically calculated as sum scores across five questions in each subdomain, each question rated on a 1–5 Likert scale; hence, sum scores range from 5 to 25, and mean subdomain sum scores range from 1 to 5. Mean scores in each subdomain were 1.72 (±0.12) for emotional abuse, 1.75 (±0.12) for physical abuse, 1.25 (±0.08) for sexual abuse, 2.70 (±0.13) for emotional neglect, and 2.32 (±0.12) for physical neglect. These scores were “low to moderate” with respect to trauma severity based on validated threshold cutoffs in each subdomain (mean cutoff, emotional abuse ≥ 1.8, physical abuse ≥ 1.6, sexual abuse ≥ 1.2, emotional neglect ≥ 2, and physical neglect ≥ 1.6)36. Subdomain mean scores were highly correlated within each domain for abuse (emotional vs. physical r = 0.64, p < 0.0001, emotional vs. sexual r = 0.34, p = 0.02, and physical vs. sexual, r = 0.36, p = 0.02) and for neglect (emotional vs. physical r = 0.5, p < 0.0001), as well as across abuse and neglect domains (r = 0.65, p < 0.0001). However, the severity of neglect was significantly greater than severity of abuse (p < 0.0001) in our sample. Henceforth, we use childhood neglect as the specific subdomain variable of interest. Participants did not have any comorbid psychiatric conditions or drug abuse history as assessed in clinical interviews.
We conducted multidimensional assessments in all study participants, including neuroimaging as the primary assessment, and complementary secondary assessments, i.e., computerized objective cognitive assessments, caregiver-based ADHD behavioral ratings, and teacher-based academic performance ratings. Neuroimaging and cognitive assessments were performed at two time points 8 weeks apart (time 1 and time 2), interspersed by the intervention period. Behavior ratings were performed at time 1, time 2, and additionally at a 1-year follow-up, time 3. Academic performance ratings could not be obtained at time 1, but were obtained at time 2 and time 3. The overall study design is depicted in Fig. 1.
Resting-state functional magnetic resonance imaging (rs-fMRI) provides a measure of spontaneous, intrinsic brain activity that can be used to assess multiple functional brain networks37,38. We implemented rs-fMRI in this study as it is more feasible in this pediatric population, and also produces larger, more robust, and reliable brain signals of energy consumption than task-based approaches39,40,41,42. Scans were conducted at AIIMS New Delhi, on a 3.0 T MR Scanner (Phillips Ingenia) equipped with a 32-ch head coil. Scans were performed in 44 of the 45 study participants; one participant was excluded due to dental braces that produce large artifacts during scanning. Anatomical T1-weighted images were collected using a high-resolution 3D magnetization-prepared rapid gradient echo sequence with 360 1-mm-thick sagittal slices (echo time [TE] 3.7 ms; repetition time [TR] 8.1 ms; field of view [FOV] 240 mm; flip angle 8°). Rs-fMRI images were acquired at rest with eyes open, gaze at a central fixation cross, using single-shot echo-planar T2*-weighted imaging sequence. Each volume consisted of 35 contiguous 4-mm-thick slices with ascending slice order and no interslice gap (TE 30 ms; TR 2000 ms; FOV 230 mm; flip angle 90°; duration 6.83 min). The scan duration we used was optimized for obtaining a sufficient number of reliable scans in pediatric populations43,44,45, which consisted of 200 volumes plus 5 initial unscored dummy volumes acquired in 6.83 min.
Rs-fMRI data were preprocessed in SPM12 (The Wellcome Department of Cognitive Neurology, London, UK, http://www.fil.ion.ucl.ac.uk/spm/software/spm12/) using standard spatial preprocessing steps. Functional data were slice-time corrected, realigned to the first image of the resting scan, normalized in Montreal Neurological Institute (MNI) space, and smoothed with a 6-mm kernel (full width at half maximum). Functional connectivity analysis was performed using a seed-driven approach using the CONN toolbox v17 (http://www.nitrc.org/projects/conn)46.
Physiological and other spurious sources of noise were estimated and regressed out using the anatomical CompCor method (aCompCor) that has been shown to yield higher specificity and sensitivity compared with global signal regression47,48. A temporal band-pass filter of 0.008–0.09 Hz was applied simultaneously to all regressors in the model. Residual head motion parameters (three rotation and three translation parameters plus another six parameters representing their first-order temporal derivatives) were regressed out. Artifact/outlier scans (average intensity deviating more than three standard deviations from the mean intensity in the session or composite head movement exceeding 1 mm from the previous image) were also regressed out to minimize the spurious effects induced by motion artifacts49. Outlier images were modeled as nuisance covariates. Each outlier image was represented by a single regressor in the general linear model (GLM), with a 1 for the outlier timepoint and 0 elsewhere. We confirmed that head displacement for either frame-to-frame translations or rotations did not significantly differ across time 1 and 2 scans (mean ± standard error of xyz translations, time 1: 0.05 ± 0.007 mm, time 2: 0.05 ± 0.004 mm; rotations, time 1: 7 × 10−4 ± 2 × 10−4 radians, time 2: 6 × 10−4 ± 8 × 10−5 radians). The number of outliers also did not significantly differ between time 1 and 2 scans (time 1: 19 ± 3, time 2: 18 ± 3). We additionally confirmed that when participants are partitioned into intervention groups, that there are no significant differences in head displacement parameters or outlier scans between intervention groups (p > 0.5), nor any intervention group × time interaction (p > 0.25).
We analyzed the dACC seed region connectivity to both cingulo-opercular and frontoparietal network regions of interest (ROIs) as dACC is found to be functionally connected to several of these regions during development11,12,13,14,15,16. The dACC seed region was defined as a 12-mm radius sphere around peak coordinates (MNI x, y, z: −2, 7, 50); 11 ROIs were similarly specified in the frontoparietal network (right/left precuneus, mid cingulate, right/left dorsolateral prefrontal cortex, right/left frontal cortex, right/left inferior parietal lobule, and right/left intraparietal sulcus) and 6 ROIs specified in the cingulo-opercular network (right/left anterior insula/frontal operculum (aI/FO), right/left anterior thalamus, and right/left anterior prefrontal cortex) as per previously described ROI coordinates11,31. Time series of all voxels within each ROI were averaged, and first-level correlation maps were produced by extracting the residual BOLD signal time course from the dACC seed ROI and computing Pearson correlation coefficients between its time course and the time course of all other ROIs. Correlation coefficients were converted to normally distributed Z scores using the Fisher transformation to allow for second-level GLM analyses. Mean functional connectivity strengths of the dACC to the frontoparietal network and to the cingulo-opercular network were calculated as the mean functional connectivity between the dACC seed and the 11 frontoparietal network ROIs and the 6 cingulo-opercular network ROIs, respectively.
In addition, we performed seed–voxel correlations by estimating maps showing temporal correlations between the BOLD signal from the dACC seed region and the time course of all other voxels. Pearson correlation coefficients were converted to normally distributed Z scores using the Fisher transformation to allow for second-level GLM analyses. For this seed–voxel connectivity data, cluster-level threshold was set at p < 0.05 using false-discovery rate correction for multiple comparisons, with voxelwise threshold of p < 0.0150.
Objective cognitive assessments
We tested participants on two standard objective cognitive assessments that are frequently tested in children with ADHD, evaluating (a) sustained attention to goal-relevant information, modeled after the test of variables of attention51, and (b) interference resolution as per the Flanker test52. In the sustained attention test, participants detected sparse targets (black square in the visual upper field appearing on 33% of trials) and withheld responses on frequent nontargets (black square in the visual lower field appearing on 67% of trials). Accuracies on this task are typically at ceiling, and the relevant response measure is the response time variability, with lower response variability indicative of higher consistency and better performance53. In the Flanker test for interference resolution, participants viewed an array of five letters and identified the central target letter while ignoring the flanking distractor letters. The letters, a/b/c/d were used, with each serving as a target or flanking letter on an equivalent number of trials. In all, 50% of task trials were congruent with matching targets and distractors, and 50% were incongruent with different targets and distractors. The main performance measure on this task is the response time cost (i.e., RT on incongruent trials minus RT on congruent trials), with smaller response time costs indicative of better performance.
Caregivers rated inattention and hyperactivity behaviors on the standard ADHD-RS IV rating scale54 that has also been previously used in research in India21,22. The same caregiver for each child rated ADHD behaviors at time 1–3. Raw scores, clinically normed percentiles, as well as the number of adolescents that surpassed clinical 80% threshold are summarized in Table 1.
Teachers rated academic performance on the academic performance rating scale (APRS55) at time 2, and different teachers rated performance at time 3; these ratings could not be obtained at time 1. On the APRS, the teacher rates the child’s math, reading, writing, and oral abilities, both in terms of accuracy and consistency. These ratings had some missing data (8 of 45 participants’ missing data at time 2, and 10 of 45 participants’ missing data at time 3).
After baseline assessment time 1, participants were cluster-randomized into either the IAI, EAI, or no Intervention (NI) arms. Cluster randomization was based on enrollment in pre-existing after-school groups. The 45 study participants were pre-enrolled in 6 separate after-school groups based on common gender and age, and common home area. Hence, we randomized two after-school groups each to the IAI (n = 15, 7 female, 8 male), EAI (n = 15, 6 female, 9 male), and NI (n = 15, 2 female, 13 male) intervention arms.
This sample size of 15 participants per study arm (IAI/EAI/NI) for a total sample of 45 participants was sufficiently powered to obtain a large effect size (η2 ≥ 0.1456) between-group (IAI/EAI/NI) effect on the primary outcome, i.e., change in functional connectivity of the dACC to the frontal opercular aI/FO region in the cingulo-opercular network, using a repeated measures analysis of variance (rm-ANOVA) with two repeated measures (baseline vs. post intervention), powered at 0.8 with alpha level of 0.05. A sample size by power plot for this calculation obtained using the G*Power tool57 is shown in supplementary figure S1.
Both IAI and EAI were self-administered as digital tablet apps for up to 30 min of practice per session (25 min of training interspersed with short 1-min breaks every 5 min), for 30 sessions over 30 nonconsecutive days (~6 weeks). Participants had no prior exposure to IAI/EAI, or interventions of this kind. To facilitate full adherence and troubleshoot any technical issues, a research staff member was present during all after-school group-training sessions; participants sat in a group, yet, performed their individual training sessions. As a result, all participants in the IAI and EAI arms had 100% intervention adherence. Participants in the NI arm were not provided any intervention, and went about their daily activities as usual between time 1 and time 2 assessments.
The intervention arms were double-blinded; both IAI and EAI arms were experimental; hence, neither the participants nor the research staff interacting with the participants had any knowledge as per the relative efficacy of one or the other arm. Also, the NI arm did not have knowledge of the other IAI/EAI arms, thereby, equating placebo effects in all study arms as much as possible. Caregivers had knowledge that their child was enrolled in a digital intervention, but were blind to the goals of the intervention. Teachers were also intervention-blind, i.e., without any knowledge if a child was participating in this intervention study. At the end of each intervention session, progress and performance data were automatically transferred to a secure study data server in de-identified format.
Participants in the IAI group practiced attending to the sensations of their breath, with monitoring guided using a digital app—Meditrain—which recently showed benefits on sustained attention in healthy young adults26. Participants were instructed to acknowledge internally distracting thoughts when they occurred during the practice, then disengage from the thought and shift their attention back to their breath. Participants practiced attention to breath in a closed loop, i.e., performance-adaptive trial durations starting as short as 10 s and progressively built up to several minutes of breath focus. Trial durations were adapted based on end-of-trial feedback from participants. At the end of each trial, participants were prompted to report, via a screen-tap, whether their attention remained on their breath throughout the trial, or if their attention was diverted by distracting thoughts. If they reported successful attention to their breath for the entire trial, the duration (in seconds) of the next trial was increased by 10%; if unsuccessful, the duration of the next trial was decreased by 20%. Using this adaptive algorithm, the intervention targeted the participants’ ability to self-regulate internal attention on an individualized basis. Intervention sessions were linked, such that the next session began at the level (i.e., trial duration) attained at the end of the previous session. The IAI participants predominantly reported successful attention to breath (reported percent success mean ± standard error 92.9 ± 0.96%; range 85.3–98.3%) with no significant differences in these percent reports in the first half (sessions 1–15) vs. second half (sessions 16–30) of the intervention (p > 0.8). Given the closed-loop nature of the program that lengthens the duration of the breath awareness period after each success, trial durations were significantly longer at the end relative to mid-intervention (mid-intervention trial duration: 93 ± 13 s, end-of-intervention: 793 ± 64 s; p < 0.0001).
Participants in the EAI group practiced attention to sensory (visual and auditory) stimuli amidst sensory distractors in the context of five different game modules, practiced 5 min each per session. All EAI modules were closed loop, i.e., performance-adaptive and adjusted task difficulty so that participants maintained ~80% performance accuracy at all times. Game modules challenged focused attention, divided selective attention, as well as working memory when a given sensory target had to be retained amidst varied distractors over several trials. The Freeze Frame, Double Decision, Mind’s Eye, and Target Tracker visual game modules, and the Hear Hear and Memory Grid auditory modules available at brainhq.com were selected for the EAI based on demonstrated efficacy of these individual training modules in prior research22,23,58,59,60. In Freeze Frame, participants practiced focused attention to visual targets, selectively withholding their response to these while non-selectively responding to all distractors. In Double Decision, participants practiced divided attention to central and peripheral visual targets. In Mind’s Eye, participants identified visual targets amidst simultaneous visual distractors as the features of the visual distractors adaptively resembled the visual target over successive trials. In Target Tracker, participants attended to moving object targets amidst moving distractors, and were adaptively challenged to retain a larger number of target objects in working memory. In Hear Hear, participants attended to target sounds within sequences of distractor sounds that adaptively resembled the target sound over successive trials. In Memory Grid, participants matched pairs of sound clips shuffled among a set of several sound clips of adaptively increasing set size.
Intervention effects in the primary neuroimaging and secondary cognitive assessments were analyzed in SPSS software using general linear modeling, specifically, group × time repeated measures ANOVAs with between-group factor of intervention (IAI, EAI, and NI) and within-group factor of assessment time (time 1 and 2). All rm-ANOVAs, including baseline covariates of childhood, neglect severity, age, gender, and age at which stable childcare access was obtained to control for differences in these variables across participants; rm-ANOVA group × time interactions were also verified that they did not differ in the absence of covariates. Estimates of effect size were reported as eta squared calculated as the ratio of the sum of squares for the effect relative to the corrected total sum of squares in SPSS (η2 < 0.06 is small, 0.06–0.14 are medium, and ≥0.14 are large effect sizes56). Post hoc testing used two-tailed paired t tests.
Intervention effects on ADHD behavioral ratings were analyzed using the nonparametric Kruskal–Wallis test, systematically investigating between-group (IAI, EAI, and NI) differences at time 1, time 2, and at the 1-year follow-up, time 3. Post hoc testing used two-sided Wilcoxon signed rank tests. The Kruskal–Wallis test was also used to investigate between-group (IAI, EAI, and NI) differences in teacher-based academic performance rating raw scores at time 2 and time 3.
For equivalent representation of results across the multidomain assessments (i.e., functional connectivity in rs-fMRI, cognitive performance, and behavioral and academic ratings), individual data were converted to Z scores within each domain. Z scores were calculated relative to the mean and standard deviation of the baseline (time 1) assessment data across all participants (n = 45) in each of the neuroimaging, cognitive, and behavioral domains. For academic data, Z scores were independently calculated at time 2 and 3, since baseline time 1 data were absent, and different teachers provided the ratings at time 2/3.