Post-mortem studies have indicated that virtually all persons with DS over age 40 harbor abnormal degrees of Aβ accumulation (7). This post-mortem evidence has since been borne out by in vivo evidence using [¹¹C]-PiB PET imaging (8, 9). In addition, several reports have highlighted the striatum as a possible nidus of amyloid accumulation, as it is commonly associated with the earliest and most prominent signal retention (8, 10), similar to that observed in a [¹¹C]-PiB PET study of presenilin-1 (PS1) mutation carriers (11). The clinical relevance of Aβ in DS has been highlighted through its associations with brain atrophy (12) and mild cognitive impairment in DS (13). While longitudinal data is scarce, Aβ accumulation in DS rates are similar to that observed in late-onset AD (14), despite an earlier onset by ~15–20 years (8, 10). More recently, tau accumulation in adults with DS has been studied using the PET tracer [¹⁸F]-AV-1451. The well-documented coupling between tau and amyloid has also led to proposals that elevated tau, particularly its propagation beyond the medial temporal cortices, is predicated upon Aβ positivity (15), consistent with that noted in late onset AD (16). Further studies to investigate the extent to which tau is associated with downstream brain atrophy (17) and cognitive decline is a subject of future research.

Glucose metabolism is one of the most robust biomarkers of AD, and [¹⁸F]-FDG PET is widely recognized as a sensitive tool to measure neuronal activity on glucose metabolism. In DS, the spatial distribution of hypometabolism appears to resemble that of sporadic AD, encompassing key regions such as the posterior cingulate and other regions that are key nodes of the default mode network (18, 19). Decreased metabolic activity has also been related to elevated Aβ burden and cognitive function (9, 20–23). Recently, the relative tracer delivery (R₁) from [¹¹C]-PiB PET imaging has been validated by Mak et al. (24) as a surrogate index of cerebral perfusion, showing marked deficits that were associated with amyloid deposition and longitudinal cognitive decline.

The cortical signature of DS has been a subject of extensive investigations using T₁-weighted MRI (9, 12, 19, 25). A common finding in the literature points is that of Aβ-associated atrophy in temporo-parietal cortices and subcortical atrophy, hallmarks of atrophy patterns in sporadic AD (26). Interestingly, there are also several reports of increased cortical thickness in DS individuals without amyloid burden (25). Taken together, atrophy appears to be predicated on the presence of amyloid, a notion that is supported by a previous study demonstrating a correlation between Aβ and atrophy in DS (15).

Diffusion tensor imaging (DTI) is a non-invasive in vivo technique for characterizing the microstructural properties of white matter tracts by quantifying changes in both the rate and directionality of water molecules (27). However, this topic has received relatively little attention compared to gray matter analyses. One study showed that persons with DS and dementia have decreased white matter integrity compared to non-demented DS subjects in a similar topography seen in AD (28, 29). Given that the role of amyloid and tau accumulation underpinning white matter dysfunction is still unclear, our group is pursuing this topic via a joint-analysis of [¹¹C]-PiB, [¹⁸F]-AV1451 and DTI datasets.

The intersection of the clinical understanding of brain pathology and technological innovation in neuroimaging has yielded significant progress in the fields of psychiatry and neurology. However, single modality imaging is inherently limited in the information it can reveal, restricting the understanding of complex and multifaceted disease processes. The complexity problem of neuroimaging datasets in neurological study and the dynamic nature of pathology is difficult to address. The successful and meaningful integration, therefore, of multiple brain imaging modalities is key to clinical progress, whether that be rooted in biomarker identification or translation to therapeutics.

Machine learning is emerging as a significantly useful technique which can both reduce and utilize the high dimensionality of neuroimaging data (30). Thus, supervised machine learning approaches are excellent candidates for the integration of multiple imaging modalities and data types. The basis of machine learning is that a data-driven methodology is applied to predictively model either a discrete or continuous outcome variable. For instance, a well-performing model based on medically non-invasive data, such as structural or functional MRI, which effectively predicts a variable only identifiable to date in an invasive or costly manner, such as amyloid burden in Alzheimer's disease, is of considerable use in clinical practice. In Down's syndrome, such advances are particularly warranted given the need for careful ethical considerations of patient vulnerability and burden. It should be noted however, that unbiased and large datasets are required for such machine learning approaches, as smaller data with sample bias may risk models that are ungeneralizable to the wider population.

Supervised machine learning algorithms are commonly applied using either a classification or regression approach, where the outcome variable of interest is either binary or continuous, respectively (31, 32). These methodologies allow for a tailored hypothesis-driven approach to the data. In Down's syndrome, an important example of the utility of a classifier would be using brain MRI data to define a patient as positive or negative for amyloid burden, thus removing the need for expensive and potentially distressing PET scanning. Given that amyloid positive status is an early indicator of the definitive development of Alzheimer's pathology, this methodology may be utilized as a biomarker for the indication of need for therapeutic intervention. Similarly, the training of regression machine learning algorithms on multimodal brain imaging data may infer accurate predictions of CSF-based biomarkers without the need for invasive procedures and laboratory cost. Importantly, machine learning represents a significantly useful technique for the integration of different types of brain imaging modalities, which among other benefits, such as predictive power, provides a wider and more accurate insight into brain pathology, leveraging multiple data types meaningful to underlying biology (33). Additionally, such implementation of machine learning for multimodal neuroimaging datasets would allow for the assessment of models for predicting pertinent clinical states and, where longitudinal training data is available, prognostic value. The limitations of machine learning techniques however should also be acknowledged, such as to-date minimal uptake of published algorithmic diagnostic approaches into clinical communities and the necessity of large and generalizable datasets for training. DS-AD is a promising candidate for such methodology as on the whole, very high disease penetrance means that the clinical population is relatively homogenous compared to other psychiatric disorders and dementias. To address the issue of translation into the clinic, neuroimaging research employing machine learning tools should carefully consider both the representative nature of the dataset and test hypotheses that address significant clinical need.

Additionally, when considering brain structure and connectivity, statistically integrating multiple MRI techniques can also be advantageous in understanding fundamental biological mechanisms. Two of the most commonly acquired types of structural data are T1-weighted and diffusion-weighted images, which rely on the time taken for the proton spin to realign with the static B₀ magnetic field and the anisotropy of water diffusion along multiple magnetic gradient b-vectors, respectively (34). Therefore, the combination of T1-weighted and diffusion-weighted imaging allows for tissue macrostructure and the tissue integrity to be analyzed in tandem. One example of such an approach is the using statistical covariance of regional brain structural metrics and diffusivity metrics like fractional anisotropy to produce a morphometry similarity index, which has been previously used to demonstrate significant underlying pathology in disorders such as psychosis (35). Application of such integration between structural modalities gives enhanced understanding of alterations to brain tissue during disease states and would be particularly useful in mapping AD development in Down's syndrome.

When considering PET modalities, atlas-based parcellations of T₁-weighted structural images are routinely used to label anatomical brain regions for the extraction of degree of tracer binding. However, other modalities may be leveraged in this approach to assess neuropathology in tandem with other metrics, such as functional connectivity, as demonstrated by Franzmeier et al. (36). Through the application of linear regression, with the vectorized functional connectivity matrix serving as a predictor for degree of longitudinal tau change as measured by AV1451, the results showed a functional connectivity and tau-spread coupling, which supports the hypothesis of trans-neuronal tau propagation in sporadic AD. Hence, the multimodal approach to data analysis in this case yielded important results for understanding the pathological development mechanisms of AD. In the case of DS-AD, connectomic approaches may be used in a similar way, in tandem with PET-based neuropathology data, for example to assess tau propagation with respect to white matter structural density. The outline of such a methodological approach is summarized in Figure 1.

The increasingly common usage of ultra-high field MRI in clinical research is also a valuable asset for the advancement of methodologies that integrate multiple brain imaging modalities. The enhanced signal-to-noise ratio and spatial resolution of high field scanners, along with the application of Bayesian probabilistic approaches to structural data has allowed for the automated segmentation of small brain structures with previously unattainable accuracy (37). The hippocampal subfields (dentate gyrus, cornu ammonis 1–4 and subiculum) can now be isolated and volumetrically analyzed from T₁- and T₂-weighted acquisitions (38, 39), with the potential to provide valuable insight in Down's syndrome given the prominent role of the hippocampus in Alzheimer's neuropathological development. However, to gain a fuller understanding of how AD development affects the hippocampal subfields, a multimodal approach should be leveraged. The treatment of the subfields as regions of interest for tau or amyloid PET analysis would yield significant insight if utilized cross-sectionally across age groups, enabling trajectory-mapping of AD pathology within the hippocampal structure. Additionally, the integration of the hippocampal subfields with structural or functional brain connectivity data as seed regions of interests would progress past basic whole brain analyses and demonstrate hippocampal-specific changes in connectivity and its relevance to clinical phenotypic expression that may further understanding of the pathological role of these granular regions during the development of DS-AD.

The statistical joint analysis of multi-modal imaging data is also a subject of future interest (40). The recently proposed non-parametric combination (NPC) technique could be employed to investigate how multiple imaging measurements are simultaneously associated with dementia or amyloid burden in DS (i.e., “Is amyloid burden associated with both gray matter and perfusion deficits in the same regions?”). For instance, it has been shown that NPC could yield additional information about patterns of group differences not visible in each modality alone (41). To the best of our knowledge, no study has applied this technique to investigate joint changes in multiple modalities and thus should be an area of future research, particularly in the early stages such as DS persons who are cognitively normal.