Individuals were considered for inclusion if they were 50–85 years old. Fifty-nine participants suspected of ADHD were identified among patients presenting to a cognitive neurology clinic. All were screened using the Adult ADHD Self-Report Scale (ASRS), and a subset (n = 37) also agreed to complete the ADHD module of the Structured Clinical Interview for DSM-5 (SCID-5). The remaining 22 were unavailable or refused to complete the SCID-5.
Participants were classified as ‘ADHD’ if they met SCID-5 criteria for adult ADHD (n = 22), with the exception of the age-of-onset criterion (< age 12). Efforts were made to ascertain that symptoms had been longstanding; however, the age-of-onset criterion was not strictly applied due to concerns with relying on very remote retrospective recall (consistent with expert recommendations to exercise some flexibility in applying this criterion in adults(19, 20)). When SCID-5 had not been completed, ASRS scores were considered: 16 participants endorsed ≥ 4 items above a pre-specified threshold value on Part A or obtained a total summed score of ≥ 14 on Part A and were considered to have significant ADHD symptoms.(21) Thirteen suspected ADHD participants did not meet SCID-5 criteria, and six had no SCID-5 data available and screened negative on the ASRS; these 19 were excluded. Two participants reported longstanding issues with attention and executive functions and obtained ASRS ≥ 14; although they did not meet SCID-5 criteria at the time of enrollment because therapy had successfully helped them manage the functional impact of these symptoms, they were assumed to have underlying ADHD and were included in the final ADHD sample.
The Conners Adult ADHD Rating Scale (CAARS) Self-Report Long Form was administered to quantify ADHD symptom severity. Participants with children were also queried about their children’s early-life ADHD symptoms using the Barkley Adult ADHD Rating Scale-IV (BAARS-IV) Other-Report Childhood Scale. This served as additional corroborative evidence of ADHD, as heritability studies estimate that ADHD occurs in 40–50% of parents who have a diagnosed child.(22) The CAARS and BAARS scores were not used to exclude any participants. Three participants were taking stimulant medication and 32 were not. Current medication was unavailable for 5 ADHD participants.
Twenty-four MCI participants were drawn from the Sunnybrook Dementia Study (SDS) (ClinicalTrials.gov NCT01800214), a well-characterized cohort with varying cognitive impairment due to neurodegenerative or vascular disease. SDS participants are thoroughly screened to exclude secondary causes of impairment or concomitant illness; thus, any suspected cases of ADHD were not included, and participants were free from stroke. SDS diagnoses are determined by at least two experienced clinicians, based on neurological and cognitive examination. Data from an additional six MCI participants enrolled in another study were also used here, bringing the total MCI sample to n = 30. All were diagnosed based on Petersen’s(23) criteria (subjective and objective impairment in any cognitive domain, preserved functional independence, and no dementia). Based on Jak and Bondi’s comprehensive neuropsychological criteria,(24) three (10.3%) had single-domain amnestic MCI, 14 (48.3%) had multiple-domain amnestic MCI, two (6.9%) had single-domain non-amnestic MCI, and eight (27.6%) had multiple-domain non-amnestic MCI. Two had unclear neuropsychological profiles based on the Jak and Bondi criteria, but evidenced isolated impairments in California Verbal Learning Test (CVLT) learning and recognition and Stroop color-naming, or Wisconsin Card Sorting Test (WCST) set-loss errors. MCI participants who could be contacted (n = 17) also completed the ASRS, BAARS and CAARS.
Although cognitive tests were used to determine group status in the MCI group and as primary outcome measures in this study, their use in combination with ADHD symptom scales aimed to decrease circularity. Participants with cognitive impairment were not necessarily all classified as MCI; those who also screened positive on the SCID or the ASRS were classified as ADHD. ADHD classification was made agnostic to cognitive status.
An additional sample of 37 healthy controls was selected from the SDS. These participants reported no cognitive complaints, performed within normal limits on all cognitive measures, and were stroke-free. The final sample consisted of 40 participants with ADHD, 29 with MCI and 37 healthy controls.
Cognitive and behavioral measures. Participants completed the Mini-Mental State Examination (MMSE) as an estimate of global cognitive function. To quantify depressive symptoms, participants completed either the 30-item Geriatric Depression Scale (GDS) (n = 85), the Beck Depression Inventory II (BDI-II) (n = 17) or an informant completed the Cornell Scale for Depression in Dementia (n = 4). In four ADHD participants, the MMSE was completed > 1 year after or prior to the rest of the neuropsychological assessment and was therefore coded ‘missing’. One ADHD and two control participants did not complete any depressive symptom measure.
All participants underwent neuropsychological assessment, but because data were drawn from different sources not everyone completed all tests. Table 2 summarizes the number of participants having completed each test. Domains assessed included attention (forward digit span, Trails A, digit-symbol coding, Stroop word-reading and color-naming), episodic memory (Logical Memory Short Story, CVLT, Rey-Osterrieth Complex Figure Task [ROCFT]), language (Boston Naming Test [BNT], phonemic and semantic fluency), and executive abilities (WCST, backward digit span, Stroop interference). Total time obtained on Trails B was transformed to a B/A ratio to isolate a relative measure of switching.(25)
Neuroimaging measures. A subset of MCI (n = 21) and controls (n = 33) had usable brain magnetic resonance imaging (MRI) data collected as part of SDS, and imaging data were acquired on an additional subset of 26 ADHD participants using the same imaging protocol. The MRI protocol was acquired on a 1.5 T GE Signa scanner (Milwaukee, WI, USA) and included a T1-weighted axial three-dimensional spoiled gradient recalled echo (5 ms echo time [TE], 35 ms repetition time [TR], 1 number of excitations [NEX], 35° flip angle [FOV], 22 × 16.5 cm, 0.859 × 0.859 mm in-plane resolution, with 1.2—1.4 mm slice thickness depending on head size) and interleaved PD and T2 sequences (interleaved axial dual-echo spin echo: TEs of 30 and 80 ms, 3 s TR, 0.5 NEX, 20 × 20 cm FOV, 0.781 × 0.781 mm in-plane resolution, 3 mm slice thickness).
Cortical thickness analysis (described previously)(26, 27) was conducted using an enhanced modification of FreeSurfer software v.6.0 (http://surfer.nmr.mgh.harvard.edu/). Briefly, the pre-processing of T1-weighted scans included motion correction, skull-stripping, transformation to Talairach space, intensity normalization, hemispheric separation, and tissue segmentation and parcellation. We performed two additional stages to the conventional pipeline to improve accuracy and quality assessment based on the PD/T2 sequence which shows the subarachnoid enabling more accurate delineation of the cortical surface. Stage 1 involved replacing the skull-stripped brain in FreeSurfer with one generated using our in-house semi-automatic brain extraction pipeline (SABRE),(28) which enhances the overall downstream processes in FreeSurfer. This was done Stage 2 involved incorporating lesions masks from our in-house PD and T2 based lesion segmentation pipeline to account for small vessel disease such as white matter hyperintensities.(29, 30) Grey and white matter, and grey matter and cerebrospinal fluid (CSF) borders, were identified and modelled as surfaces. Cortical thickness was defined as the distance between the grey and white matter surface boundaries and the grey and CSF boundaries along each point of the cortex in each hemisphere. After pre-processing, all participants’ surface data were resampled to FreeSurfer’s average surface map. A 15-mm full-width half-maximum Gaussian spatial smoothing kernel was applied to the surface maps. FreeSurfer outputs, based on the Desikan atlas parcellation,(31) were visually inspected for quality control.
T1 was segmented using a multi-feature histogram method to generate a tissue segmentation containing normal appearing grey matter (NAGM), normal appearing white matter (NAWM), sulcal and ventricular CSF.(32) SABRE was used to parcellate brain tissue into 26 standardized volumes of interest described elsewhere.(28) Hippocampal volumes were segmented using an in-house 3D convolutional neural network with a U-net architecture that is robust for populations with brain atrophy (https://hippmapp3r.readthedocs.io).(33)
Cognitive and behavioral measures. Depressive symptoms were categorized as ‘none’ (GDS < 10; BDI-II < 14; Cornell < 8), ‘mild/probable’ (GDS 10–19; BDI-II 14–19; Cornell 8–12) or ‘moderate/severe’ (GDS > 19; BDI-II > 19; Cornell > 12). Age, education, and MMSE scores were non-normally distributed, therefore nonparametric Kruskal-Wallis H tests were applied. Sex and depression were compared using chi-square (χ2). ASRS, CAARS and BAARS scores were compared using t-tests.
Raw cognitive scores were standardized to Z scores using published normative data. Z scores were entered into separate univariate analysis of variance models adjusted for age, sex, MMSE and depressive symptoms (because the groups differed on these measures; see Results below). Models were then adjusted to remove non-significant predictors except age. Pairwise comparisons were examined where the main effect was significant.
Neuroimaging measures. NAGM volumes in inferior, middle, and superior frontal regions, superior and inferior parietal regions, and anterior and posterior temporal regions were corrected for head size by dividing cubic millimeters in each region by total supratentorial intracranial volume. Corrected regional volumes were normally distributed and entered raw into univariate analysis of variance models adjusted for age, sex, MMSE and depressive symptoms. Models were individually adjusted to remove non-significant predictors except age. Pairwise comparisons were examined in models with a significant main effect.
For cortical thickness analyses, vertex-wise surface-based analysis was first performed within the frontal lobe only using the general linear model in FreeSurfer, based on cortical thickness alterations in young adults with ADHD(34). A second exploratory analysis was performed across the whole brain. Age, education, MMSE scores were included as regressors of no interest. Monte Carlo simulation with 5000 iterations using a cluster-wise probability (p(cwp)) of p < 0.05 (two-sided) was used to correct for multiple comparisons. Bonferroni correction was applied across the two hemispheres.