The Irish Longitudinal Study on Ageing (TILDA) is a large population-based study of a nationally-representative sample of community-dwelling older adults aged 50 years and over, resident in the Republic of Ireland. TILDA’s random sampling procedure and study design have been described elsewhere (Whelan and Savva, 2013). Briefly, participants have been followed biennially from Wave 1 (data collection baseline: 2009–2011; N = 8,504) and this analysis includes data from Wave 3 (2014–2015), Wave 4 (2016–2016), Wave 5 (2018–2018) and Wave 6 (2021–2021). Macro- and micro-structural measures of WMHs at Wave 3 and cognitive data (Wave 3–6) were available for 497 participants (see Figure 1 for a detailed breakdown). At each wave, participants took part in a computer assisted personal interview (CAPI) and completed a questionnaire (SCQ). At Wave 3, participants were also invited to take part in a health assessment and a sub sample of those who completed the health assessment were randomly selected to undergo magnetic resonance imaging (MRI). Ethical approval was granted by the Trinity College Faculty of Health Sciences Research Ethics Committee, Dublin, Ireland. Protocols conformed with the 1964 Declaration of Helsinki and its later amendments. Signed informed consent was obtained from all respondents prior to participation. Additional ethics approval was received for the MRI sub-study from the St James’s Hospital/Adelaide and Meath Hospital, Inc. National Children’s Hospital, Tallaght (SJH/AMNCH) Research Ethic Committee, Dublin, Ireland. Those attending for MRI were also required to complete an additional MRI-specific consent form.

The MRI sampling procedure has been previously described in detail (De Looze et al., 2020). In brief, MRI scanning was completed at the Thomas Mitchell Centre for Advanced Medical Imaging (CAMI), St. James’ Hospital, Dublin, (3 T Philip’s Achieva system and 32-channel head coil). The study protocol included T1, fluid attenuation inversion recovery (FLAIR) and diffusion weighted imaging (DWI) sequences. T1-weighted scans were acquired using a 3D Magnetisation Prepared Rapid Gradient Echo (MP – RAGE) sequence (FOV (mm): 240 × 240 × 162; SENSE factor: 2; TR: 6.7 ms; TE: 3.1 ms; flip angle: 8°; voxel size (mm): 0.8 × 0.8 × 0.9). FLAIR images were acquired using a 2D turbo spin-echo sequence (FOV (mm): 230 × 230; TR: 11000 ms; TE: 125 ms; TI: 2800 ms; flip angle: 90°; number of slices: 30; slice thickness: 4 mm; slice gap: 1 mm; voxel size(mm): 0.45 × 0.45). DWI images were acquired using a high spatial (1.9 mm³) and high angular resolution (61-direction) protocol with a high diffusion sensitivity (b = 1,200 s/mm²; TR: 2885 ms; TE: 64 ms; flip angle: 90°; number of slices: 30; slice thickness: 4 mm; slice gap 1 mm: voxel size (mm): 0.96 × 0.96).

Images were pre-processed using SPM12 (University College London, London, UK) and ExploreDTI¹ (Leemans et al., 2009). All images were reorientated to a canonical SPM template. The dataset was inspected for appropriate orientation and any gross artifacts before processing and segmentation. DWI preparation included correcting multiple prerequisite parameters in the ExploreDTI software toolbox for MATLAB.² The entire standardised dataset was converted to ExploreDTI compatible diffusion files, corrected for signal drift (Vos et al., 2017), Gibbs Ringing artefact (Perrone et al., 2015), head motion and eddy current artefact with registration to the B0 reference image (Soares et al., 2013) and EPI deformation correction via registration to the T1 sequence (Irfanoglu et al., 2012). These commands were performed using the native ExploreDTI toolbox plugins.

White matter hyperintensities (WMHs) were segmented utilising the Lesion Prediction Algorithm (LPA) (Schmidt, 2016), implemented in the LST toolbox version 3.0.0 for SPM12.³ The LPA was applied to both FLAIR and T1 images. Meticulous visual image review ensured accurate and appropriate segmentation of only WMHs. Each segmented lesion received a unique lesion number, and an individualised atlas of lesions was created for each subject. To objectively localise each lesion, the SPM Anatomic Automatic Labeling extension was employed (Rolls et al., 2015). Lesions were categorised based on their anatomical position in the periventricular or the deep white matter. Periventricular lesions were defined as those within 15 mm of a ventricle (Nyquist et al., 2015), while deep white matter lesions were those outside this range. Atlas-based tractography was performed using the segmented lesions as regions of interest in each subject, utilising the native ExploreDTI plugin, and diffusion metrics (FA and MD) unique to each WMH were calculated.

Lesions were initially clustered based on both macrostructural properties (volume in mm³) and microstructural properties (FA and MD) using k-means clustering (cluster package in R version 4.0.3 with RStudio Version 1.4.1103). The optimal number of lesion phenotypes or clusters was determined using the elbow method. Three primary lesion phenotypes were identified. These phenotypes were further stratified based on their location (deep vs. periventricular), resulting in a total of six lesion phenotypes. The frequency of each lesion phenotype per individual was calculated and used for subsequent analysis.

Global cognitive function at each wave [Wave 3 to 6; MEAN ± SD-year follow-up; wave 3–28.8 (1.5), wave 4–28.6 (1.3) wave 5–27.9 (1.7) and wave 6–27.6 (1.8)] was assessed using a battery of neuropsychological tests, including 10-word immediate recall test (2 trials), 10-word delayed recall test, semantic verbal fluency (animal naming), and orientation (date, day, month, year). A composite score per participant and at each wave was derived by combining individual test scores to broadly reflect global cognitive function. The composite scores were generated based on an equally unit-weighted approach using standardised scores (z scores). The lower the composite score, the worse the cognitive function.

Covariates included age, sex, education (none/primary, secondary, or tertiary/higher); smoking (never, past or current); problematic alcohol consumption (CAGE questionnaire; yes or no); physical activity (International Physical Activity Questionnaire – IPAQ; short-form; low, moderate or high); body mass index (BMI, continuous variable); pre-existing self-reported physician-diagnosed cardiovascular diseases and events (CVDEs, none or one or more of the following: angina, heart attack, congestive heart failure, stroke, transient ischemic attack, and atrial fibrillation); average systolic and diastolic blood pressure measurements obtained twice using an OMRON digital automatic blood pressure monitor (Model M10-IT); antihypertensive medication [Anatomical Therapeutic Chemical (ATC) codes C02 (antiadrenergic agents), C03 (diuretics), C07 (β blockers), C08 (calcium-channel blockers), and C09 (angiotensin- converting enzyme inhibitors)]; antidepressant medication (ATC code N06A); and symptoms of depression (the Centre for Epidemiological Studies Depression Scale – CESD; continuous variable).

All statistical analyses were performed using R version 4.0.3 with RStudio Version 1.4.1103. The observed sample was first characterised using unadjusted means and standard deviations for continuous variables and percentages (%) for categorical variables. To estimate selection bias, the observed sample was compared to the total cohort including those excluded from analysis. Independent t-tests and chi-squared tests where appropriate were used to assess differences between the two groups. Linear mixed effect models were used to examine the relationship between lesion phenotypes (predictor) and cognitive function at waves 3, 4, 5 and 6 (outcome). Fixed effects included the number of lesion phenotypes in separate models and waves with an interaction term. The significance of the interaction was assessed using likelihood ratio tests. Participants constituted the random intercept. Baseline models included age, sex and education as covariates; further adjustment included smoking, alcohol consumption, physical activity, BMI, CVDEs, average systolic and diastolic blood pressure, depressive symptoms and antihypertensive and antidepressant medication.

Table 1 provides descriptive statistics of the study sample (n = 497) at Wave 3. Mean age of the study sample was 68.5 years (SD 7.56); 51.9% of the participants were female and 20.7% had a low (primary school) level of education. Mean MMSE (Mini Mental State Examination) at Wave 3 was 28.8 (SD = 1.5).

A total of 11,933 individual white matter lesions were segmented and analysed from the cohort (n = 497) (Table 2). On average, each participant exhibited 24 white matter hyperintensities (WMHs) with a mean volume of 263.21 mm³.

Cluster analyses revealed three distinct lesion phenotypes (Figure 2): Type A (Low Volume, High FA/Low MD), Type B (Low Volume, Low FA/High MD) and Type C (High Volume, Low FA/High MD), and, each further categorised into periventricular and deep white matter lesions. We will omit the MD descriptor for simplicity from this point. 40% of lesions were Periventricular Type A (Low Volume, High FA), 26% were Periventricular Type B (Low Volume, Low FA), 19% were Periventricular Type C (High Volume, Low FA), 11% were Deep Type A (Low Volume, High FA), 3% were Deep Type C (High Volume, Low FA), and 1% were Deep Type B (Low Volume, Low FA). Descriptive characteristics of these six lesion phenotypes are presented in Table 3.

Comparisons among the lesion phenotypes, with the Periventricular low volume, high FA Type A (Low Volume, High FA) group as the reference, revealed notable demographic and clinical differences. Those with deep phenotypes (Type A, B, C) were older (p < 0.001), more likely to be women (Type A (Low Volume, High FA) and C (High Volume, Low FA); p < 0.001), and exhibited lower levels of education (Type A and C; p < 0.001). Individuals in the Deep Phenotype groups were more likely to have two or more cardiovascular conditions (Type A (Low Volume, High FA) and C (High Volume, Low FA); p < 0.001, Type B (Low Volume, Low FA) p < 0.05), higher systolic blood pressure (Type A (Low Volume, High FA) p < 0.05, Type B (Low Volume, Low FA) p < 0.001), current smoking habits (Type A (Low Volume, High FA); p < 0.001, Type C (High Volume, Low FA); p < 0.01), problematic alcohol consumption (Type A (Low Volume, High FA); p < 0.01, Type B (Low Volume, Low FA); p < 0.001), lower levels of physical activity (Type A (Low Volume, High FA); p < 0.001, Type C (High Volume, Low FA); p < 0.05), and higher levels of depressive symptoms (Type C (High Volume, Low FA); p < 0.01, Type A (Low Volume, High FA); p < 0.1). Similar trends were observed in the Periventricular Type B (Low Volume, Low FA) group.

Baseline Linear Mixed Effects (LME) models revealed that a higher number of deep (X = 7.3; p = 0.05) and periventricular (X = 12.9, p = 0.004) high volume, low fractional anisotropy lesion phenotypes (Type C) were associated with accelerated cognitive decline from Wave 4 to Wave 6 (Figures 3, 4), compared to the reference group with periventricular low volume, high FA (Type A) lesions. Additionally, a higher number of deep low volume, high FA (Type A) lesions were also linked to accelerated cognitive decline (X = 10.5; p = 0.01) (Figure 5).

The phenotype by Wave correlation did not reach significance for the number of periventricular low volume, high FA (Type A) lesions, deep and periventricular low volume, low FA (Type B) lesions (p > 0.05). There were no significant associations between the number of periventricular Type A (low volume, high FA) lesions (p = 0.46) or deep Type B (low volume, low FA) lesions (p = 0.48) and cognitive decline. In contrast, the number of periventricular low volume, low FA (Type B) lesions was overall positively associated with cognitive function (B = 0.04; 95% CI = 0.00, 0.09; p = 0.02) (Figure 6).

These findings remained robust to adjustments for smoking, alcohol consumption, physical activity, body mass index (BMI), cardiovascular diseases (CVDEs), average systolic and diastolic blood pressure, antihypertensive and antidepressant medication, and of depression.

Our study, while offering valuable insights into the intricate relationship between white matter hyperintensity (WMH) phenotypes and cognitive decline, is characterised by both strengths and limitations that collectively contribute to the overall integrity of our findings.

Large, Representative Sample: The inclusion of a substantial MRI sample from a nationally representative population-based study enhances the external validity of our findings, allowing for broader generalisability to diverse groups of community-dwelling older adults.

Longitudinal Design: The longitudinal nature of our study spanning 6 years provides a temporal dimension to our analyses, enabling the exploration of cognitive decline over time and capturing changes in WMH phenotypes.

Advanced Lesion Phenotyping: Employing a data-driven approach for lesion phenotyping, considering both macrostructural and microstructural properties, adds depth to our characterisation of WMH subtypes. This nuanced approach enhances our understanding of the heterogeneity within WMHs.

Cross-Sectional Nature of Lesion Phenotyping: The identification of WMH phenotypes was based on cross-sectional analyses, limiting our ability to capture dynamic changes in lesion characteristics over time. Longitudinal studies with repeated imaging sessions could provide a more comprehensive understanding of the evolving nature of WMHs.

Generalisability and Population Characteristics: Our study focused on a specific cohort of community-dwelling older adults from a nationally representative population-based study. Generalisability to other demographic groups or clinical populations may be limited, and caution is advised when extrapolating findings to broader populations.

Absence of a Control Group: The absence of a control group composed of individuals without WMHs or with low lesion burden is a limitation of our study. Inclusion of a control group would have allowed for direct comparison between individuals with varied degrees of WMH burden and those with normal appearing white matter, lending strength to the subsequent findings and improving the validity of the study. In the absence of a control group it is difficult to isolate the effects of WMHs with regards to cognitive decline from other age related factors or comorbidities that could independently affect outcomes. In addition a control group would have provided a baseline trajectory of cognitive outcomes. Future research should incorporate an appropriately matched control group to enhance the outcomes and strengthen the causal relationship between WMH and cognitive decline.

MR Acquisition Technique: Our MR acquisition protocol utilises a slice thickness of 4 mm for both the FLAIR and diffusion sequences which represents a limitation of our study. A thickness of 4 mm reduces our spatial resolution and limits segmentation of smaller white matter lesions, and may contribute to partial volume effects (mixing of signals from adjacent tissue), compromising the ability to distinguish between smaller lesions. Thinner slices (2–3 mm ideally) would provide more accurate lesion segmentation. In addition the implementation of a multi-shell sequence would allow for more detailed characterisation of white matter microstructure through utilisation of multiple B value acquisitions. This would allow for diffusion kurtosis imaging and more complex analysis of white matter microstructure, enhancing the differentiation of our described white matter phenotypes. Furthermore, multi-shell imaging would mitigate the effect of partial volume errors. Adopting higher resolution images and a multi-shell technique would result in more accurate data.

Acknowledging both strengths and limitations, our study offers a nuanced exploration of the complex relationships between WMH characteristics and cognitive decline. The comprehensive approach, from advanced lesion phenotyping to consideration of diverse factors, contributes to a balanced interpretation of our findings. While our observations provide valuable insights, the identified limitations underscore the need for continued research refinement and external validation. Striking a balance between the strengths and limitations enhances the reliability and applicability of our study, paving the way for future investigations in this evolving field.